diff --git a/cmd/imagen/compare.go b/cmd/imagen/compare.go
new file mode 100644
index 0000000..c7b5ba4
--- /dev/null
+++ b/cmd/imagen/compare.go
@@ -0,0 +1,386 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"flag"
+	"fmt"
+	"image"
+	"image/color"
+	"image/draw"
+	"image/png"
+	"io"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"time"
+
+	"golang.org/x/image/font"
+	"golang.org/x/image/font/basicfont"
+	"golang.org/x/image/math/fixed"
+
+	"mgit.msbls.de/m/ImaGen/internal/backend"
+	"mgit.msbls.de/m/ImaGen/internal/config"
+	"mgit.msbls.de/m/ImaGen/internal/output"
+	"mgit.msbls.de/m/ImaGen/internal/prompt"
+)
+
+// runCompare implements `imagen compare "<prompt>" --models a,b,c --output <dir>`.
+//
+// Each backend in --models runs sequentially against the same prompt (mRock
+// has a single GPU; parallelising would just OOM). Each generation lands as
+// a backend-suffixed file in the output dir; a contact sheet stitches them
+// together into one PNG with the backend name overlaid on each cell. A
+// sidecar JSON next to the contact sheet lists every generation with its
+// per-model metadata (latency, seed, model file, VRAM peak).
+func runCompare(ctx context.Context, args []string) error {
+	fs := flag.NewFlagSet("compare", flag.ContinueOnError)
+	var (
+		modelsCSV  string
+		size       string
+		outDir     string
+		style      string
+		negative   string
+		seed       int64
+		steps      int
+		configPath string
+		noContact  bool
+	)
+	fs.StringVar(&modelsCSV, "models", "", "comma-separated backend instance names (required)")
+	fs.StringVar(&size, "size", "1024x1024", "WxH for every backend")
+	fs.StringVar(&outDir, "output", "", "directory to write the images + contact sheet (default: ~/Pictures/imagen/compare)")
+	fs.StringVar(&style, "style", "", "style preset applied to the prompt before dispatching to each backend")
+	fs.StringVar(&negative, "negative", "", "negative prompt (forwarded to every backend that supports it)")
+	fs.Int64Var(&seed, "seed", 0, "deterministic seed for every backend (0 = each backend rolls its own)")
+	fs.IntVar(&steps, "steps", 0, "diffusion steps (0 = each backend's default)")
+	fs.StringVar(&configPath, "config", "", "config file path (default: ~/.config/imagen.yaml)")
+	fs.BoolVar(&noContact, "no-contact-sheet", false, "skip the composite PNG; only write per-backend images + sidecar")
+	fs.Usage = func() {
+		fmt.Fprintln(fs.Output(), `Usage: imagen compare "<prompt>" --models a,b,c [flags]`)
+		fs.PrintDefaults()
+	}
+	leadingPositional, flagArgs := splitLeadingPositional(args)
+	if err := fs.Parse(flagArgs); err != nil {
+		return err
+	}
+	positional := append(leadingPositional, fs.Args()...)
+	if len(positional) == 0 {
+		fs.Usage()
+		return userErr("missing prompt")
+	}
+	rawPrompt := strings.Join(positional, " ")
+	modelNames := splitCSV(modelsCSV)
+	if len(modelNames) == 0 {
+		return userErr("--models is required (comma-separated backend instance names)")
+	}
+
+	w, h, err := parseSize(size)
+	if err != nil {
+		return userErr("bad --size: %v", err)
+	}
+
+	cfg, cfgErr := config.Load(configPath)
+	if cfgErr != nil && !os.IsNotExist(cfgErr) {
+		return cfgErr
+	}
+
+	if outDir == "" {
+		home, _ := os.UserHomeDir()
+		outDir = filepath.Join(home, "Pictures", "imagen", "compare")
+	}
+	outDir = config.ExpandPath(outDir)
+
+	finalPrompt, err := prompt.Apply(rawPrompt, style)
+	if err != nil {
+		return userErr("%v", err)
+	}
+
+	runID := time.Now().Format("20060102-150405")
+	runDir := filepath.Join(outDir, runID+"-"+output.Slug(rawPrompt))
+	if err := os.MkdirAll(runDir, 0o755); err != nil {
+		return fmt.Errorf("mkdir %s: %w", runDir, err)
+	}
+
+	results := make([]compareResult, 0, len(modelNames))
+	for i, name := range modelNames {
+		fmt.Fprintf(os.Stderr, "[%d/%d] %s ...\n", i+1, len(modelNames), name)
+		res, err := generateOne(ctx, cfg, name, finalPrompt, negative, w, h, seed, steps, runDir, rawPrompt)
+		if err != nil {
+			// Don't abort the whole run on a single backend failure — record
+			// the error and continue. flexsiebels-style consumers want to
+			// see N-1 results rather than zero when one model is offline.
+			fmt.Fprintf(os.Stderr, "    failed: %v\n", err)
+			results = append(results, compareResult{Backend: name, Error: err.Error()})
+			continue
+		}
+		fmt.Fprintf(os.Stderr, "    %s (%d ms)\n", res.ImagePath, res.LatencyMs)
+		results = append(results, res)
+	}
+
+	// Sidecar JSON beside the run dir captures every attempt.
+	sidecar := filepath.Join(runDir, "compare.json")
+	if err := writeCompareSidecar(sidecar, rawPrompt, style, negative, w, h, seed, steps, results); err != nil {
+		return err
+	}
+	fmt.Fprintln(os.Stderr, "sidecar:", sidecar)
+
+	// Contact sheet stitches the successful results together. If every
+	// backend failed there's nothing to draw, so skip silently.
+	if !noContact {
+		successes := successfulResults(results)
+		if len(successes) > 0 {
+			sheet := filepath.Join(runDir, "contact-sheet.png")
+			if err := writeContactSheet(sheet, rawPrompt, successes); err != nil {
+				return fmt.Errorf("contact sheet: %w", err)
+			}
+			fmt.Println(sheet)
+		} else {
+			fmt.Fprintln(os.Stderr, "imagen compare: all backends failed; no contact sheet written")
+		}
+	}
+	return nil
+}
+
+// compareResult is one backend's output in a comparison run. Error is set
+// when Generate failed for this backend; ImagePath + Metadata are empty in
+// that case.
+type compareResult struct {
+	Backend     string         `json:"backend"`
+	ImagePath   string         `json:"image_path,omitempty"`
+	Seed        int64          `json:"seed"`
+	LatencyMs   int64          `json:"latency_ms,omitempty"`
+	Model       string         `json:"model,omitempty"`
+	VRAMUsedMiB int64          `json:"vram_used_mib,omitempty"`
+	Metadata    map[string]any `json:"metadata,omitempty"`
+	Error       string         `json:"error,omitempty"`
+}
+
+func generateOne(ctx context.Context, cfg *config.Config, name, finalPrompt, negative string, w, h int, seed int64, steps int, runDir, rawPrompt string) (compareResult, error) {
+	be, err := buildBackend(cfg, name)
+	if err != nil {
+		return compareResult{Backend: name}, err
+	}
+	attachUsageSink(be)
+
+	req := backend.Request{
+		Prompt:         finalPrompt,
+		NegativePrompt: negative,
+		Width:          w,
+		Height:         h,
+		Steps:          steps,
+		Seed:           seed,
+	}
+	res, err := be.Generate(ctx, req)
+	if err != nil {
+		return compareResult{Backend: name}, err
+	}
+	defer res.ImageReader.Close()
+
+	imgBytes, err := io.ReadAll(res.ImageReader)
+	if err != nil {
+		return compareResult{Backend: name}, fmt.Errorf("read image: %w", err)
+	}
+
+	imgPath := filepath.Join(runDir, output.Slug(rawPrompt)+"--"+output.Slug(name)+"."+extFromMime(res.MimeType))
+	if err := os.WriteFile(imgPath, imgBytes, 0o644); err != nil {
+		return compareResult{Backend: name}, fmt.Errorf("write %s: %w", imgPath, err)
+	}
+
+	cr := compareResult{
+		Backend:   name,
+		ImagePath: imgPath,
+		Seed:      seedFromMetadata(res.Metadata, seed),
+		LatencyMs: metaInt64(res.Metadata, "latency_ms"),
+		Model:     metaString(res.Metadata, "model"),
+		Metadata:  res.Metadata,
+	}
+	if v, ok := res.Metadata["vram_used_mib"].(int64); ok {
+		cr.VRAMUsedMiB = v
+	}
+	return cr, nil
+}
+
+func successfulResults(rs []compareResult) []compareResult {
+	out := make([]compareResult, 0, len(rs))
+	for _, r := range rs {
+		if r.Error == "" && r.ImagePath != "" {
+			out = append(out, r)
+		}
+	}
+	return out
+}
+
+func writeCompareSidecar(path, rawPrompt, style, negative string, w, h int, seed int64, steps int, results []compareResult) error {
+	body := map[string]any{
+		"timestamp":  time.Now().UTC().Format(time.RFC3339),
+		"prompt":     rawPrompt,
+		"style":      style,
+		"negative":   negative,
+		"width":      w,
+		"height":     h,
+		"seed":       seed,
+		"steps":      steps,
+		"results":    results,
+		"backends":   backendNames(results),
+		"successful": len(successfulResults(results)),
+		"total":      len(results),
+	}
+	data, err := json.MarshalIndent(body, "", "  ")
+	if err != nil {
+		return fmt.Errorf("marshal sidecar: %w", err)
+	}
+	return os.WriteFile(path, append(data, '\n'), 0o644)
+}
+
+func backendNames(rs []compareResult) []string {
+	out := make([]string, len(rs))
+	for i, r := range rs {
+		out[i] = r.Backend
+	}
+	sort.Strings(out)
+	return out
+}
+
+// writeContactSheet stitches a grid of (image, label) cells into one PNG.
+// Cells are sized to fit in a target width of ~2400px while keeping each
+// individual image full-resolution (no downscale) up to the column limit;
+// past that, images sit at their native size and we just lay them out.
+//
+// The grid is a simple horizontal row when N <= 4; otherwise N/2 rows of 2.
+// This is a contact sheet, not a fancy gallery — readability for side-by-
+// side eyeballing is the goal.
+func writeContactSheet(path, prompt string, results []compareResult) error {
+	if len(results) == 0 {
+		return fmt.Errorf("no successful results to lay out")
+	}
+	cells := make([]contactCell, 0, len(results))
+	for _, r := range results {
+		img, err := readPNG(r.ImagePath)
+		if err != nil {
+			return fmt.Errorf("read %s: %w", r.ImagePath, err)
+		}
+		cells = append(cells, contactCell{
+			Image:    img,
+			Label:    r.Backend,
+			SubLabel: fmt.Sprintf("%dms · seed %d", r.LatencyMs, r.Seed),
+		})
+	}
+
+	cols := len(cells)
+	if cols > 4 {
+		cols = 2
+	}
+	rows := (len(cells) + cols - 1) / cols
+
+	const labelH = 64
+	const pad = 16
+
+	cellW := cells[0].Image.Bounds().Dx()
+	cellH := cells[0].Image.Bounds().Dy()
+	for _, c := range cells {
+		if w := c.Image.Bounds().Dx(); w > cellW {
+			cellW = w
+		}
+		if h := c.Image.Bounds().Dy(); h > cellH {
+			cellH = h
+		}
+	}
+
+	totalW := cols*cellW + (cols+1)*pad
+	totalH := rows*(cellH+labelH) + (rows+1)*pad + 48 // header band
+
+	canvas := image.NewRGBA(image.Rect(0, 0, totalW, totalH))
+	draw.Draw(canvas, canvas.Bounds(), &image.Uniform{C: color.RGBA{R: 30, G: 30, B: 35, A: 255}}, image.Point{}, draw.Src)
+
+	// Header: show the truncated prompt.
+	headerText := "imagen compare — " + truncate(prompt, 100)
+	drawText(canvas, headerText, pad, 30, color.RGBA{R: 240, G: 240, B: 245, A: 255})
+
+	for i, c := range cells {
+		col := i % cols
+		row := i / cols
+		x0 := pad + col*(cellW+pad)
+		y0 := 48 + pad + row*(cellH+labelH+pad)
+		// Center the image inside the cell when smaller than the max cell size.
+		iw := c.Image.Bounds().Dx()
+		ih := c.Image.Bounds().Dy()
+		offX := (cellW - iw) / 2
+		offY := (cellH - ih) / 2
+		dstRect := image.Rect(x0+offX, y0+offY, x0+offX+iw, y0+offY+ih)
+		draw.Draw(canvas, dstRect, c.Image, c.Image.Bounds().Min, draw.Src)
+
+		// Label band underneath.
+		labelY := y0 + cellH + 20
+		drawText(canvas, c.Label, x0+8, labelY, color.RGBA{R: 250, G: 250, B: 250, A: 255})
+		drawText(canvas, c.SubLabel, x0+8, labelY+22, color.RGBA{R: 180, G: 180, B: 190, A: 255})
+	}
+
+	f, err := os.Create(path)
+	if err != nil {
+		return fmt.Errorf("create %s: %w", path, err)
+	}
+	defer f.Close()
+	return png.Encode(f, canvas)
+}
+
+type contactCell struct {
+	Image    image.Image
+	Label    string
+	SubLabel string
+}
+
+func readPNG(path string) (image.Image, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, err
+	}
+	defer f.Close()
+	img, _, err := image.Decode(f)
+	return img, err
+}
+
+func drawText(dst *image.RGBA, s string, x, y int, c color.Color) {
+	drawer := &font.Drawer{
+		Dst:  dst,
+		Src:  &image.Uniform{C: c},
+		Face: basicfont.Face7x13,
+		Dot:  fixed.Point26_6{X: fixed.I(x), Y: fixed.I(y)},
+	}
+	drawer.DrawString(s)
+}
+
+func truncate(s string, max int) string {
+	if len(s) <= max {
+		return s
+	}
+	return s[:max-1] + "…"
+}
+
+func splitCSV(s string) []string {
+	parts := strings.Split(s, ",")
+	out := make([]string, 0, len(parts))
+	for _, p := range parts {
+		p = strings.TrimSpace(p)
+		if p != "" {
+			out = append(out, p)
+		}
+	}
+	return out
+}
+
+func metaInt64(m map[string]any, key string) int64 {
+	v, ok := m[key]
+	if !ok {
+		return 0
+	}
+	switch n := v.(type) {
+	case int64:
+		return n
+	case int:
+		return int64(n)
+	case float64:
+		return int64(n)
+	}
+	return 0
+}
diff --git a/cmd/imagen/compare_test.go b/cmd/imagen/compare_test.go
new file mode 100644
index 0000000..cb9152b
--- /dev/null
+++ b/cmd/imagen/compare_test.go
@@ -0,0 +1,203 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"image/png"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// runCompareWithEnv runs the compare subcommand in a writable tmpdir, with
+// XDG_CONFIG_HOME pointing somewhere empty so no host imagen.yaml leaks in.
+func runCompareWithEnv(t *testing.T, args []string) (stderr, stdout *bytes.Buffer, runDir string, err error) {
+	t.Helper()
+	tmp := t.TempDir()
+	t.Setenv("XDG_CONFIG_HOME", filepath.Join(tmp, "no-config"))
+	t.Setenv("HOME", tmp)
+
+	out := filepath.Join(tmp, "compare")
+	// stdlib flag parsing requires flags after the leading positional. Append
+	// --output at the end so any caller-supplied flags still parse cleanly.
+	args = append(args, "--output", out)
+
+	// Capture stdout/stderr via os pipes since runCompare writes directly.
+	oldStdout := os.Stdout
+	oldStderr := os.Stderr
+	rOut, wOut, _ := os.Pipe()
+	rErr, wErr, _ := os.Pipe()
+	os.Stdout = wOut
+	os.Stderr = wErr
+	defer func() {
+		os.Stdout = oldStdout
+		os.Stderr = oldStderr
+	}()
+
+	cmdErr := runCompare(context.Background(), args)
+
+	_ = wOut.Close()
+	_ = wErr.Close()
+	stdout = &bytes.Buffer{}
+	stderr = &bytes.Buffer{}
+	_, _ = stdout.ReadFrom(rOut)
+	_, _ = stderr.ReadFrom(rErr)
+
+	entries, _ := os.ReadDir(out)
+	if len(entries) == 1 {
+		runDir = filepath.Join(out, entries[0].Name())
+	}
+	return stderr, stdout, runDir, cmdErr
+}
+
+func TestCompareHappyPathWithMockBackends(t *testing.T) {
+	// Two mock instances stand in for two different backends. mock ignores
+	// cfg so we can reuse the registered type as the instance name and skip
+	// writing imagen.yaml entirely.
+	stderr, stdout, runDir, err := runCompareWithEnv(t, []string{
+		"a cat in a fishbowl",
+		"--models", "mock,mock",
+		"--size", "64x64",
+		"--seed", "42",
+	})
+	if err != nil {
+		t.Fatalf("runCompare: %v\nstderr: %s", err, stderr.String())
+	}
+
+	if runDir == "" {
+		t.Fatal("expected a run directory under --output")
+	}
+	// Sidecar JSON
+	sidecar := filepath.Join(runDir, "compare.json")
+	data, err := os.ReadFile(sidecar)
+	if err != nil {
+		t.Fatalf("read sidecar: %v", err)
+	}
+	var body struct {
+		Prompt     string `json:"prompt"`
+		Successful int    `json:"successful"`
+		Total      int    `json:"total"`
+		Results    []struct {
+			Backend   string `json:"backend"`
+			ImagePath string `json:"image_path"`
+			Error     string `json:"error"`
+		} `json:"results"`
+	}
+	if err := json.Unmarshal(data, &body); err != nil {
+		t.Fatalf("parse sidecar: %v\n%s", err, data)
+	}
+	if body.Prompt != "a cat in a fishbowl" {
+		t.Errorf("prompt = %q", body.Prompt)
+	}
+	if body.Total != 2 || body.Successful != 2 {
+		t.Errorf("counts = %d successful / %d total", body.Successful, body.Total)
+	}
+	for _, r := range body.Results {
+		if r.Error != "" {
+			t.Errorf("backend %s errored: %s", r.Backend, r.Error)
+		}
+		if _, err := os.Stat(r.ImagePath); err != nil {
+			t.Errorf("image not on disk for %s: %v", r.Backend, err)
+		}
+	}
+
+	// Contact sheet path was printed on stdout.
+	sheet := strings.TrimSpace(stdout.String())
+	if sheet == "" {
+		t.Fatal("expected contact sheet path on stdout")
+	}
+	f, err := os.Open(sheet)
+	if err != nil {
+		t.Fatalf("open contact sheet: %v", err)
+	}
+	defer f.Close()
+	img, err := png.Decode(f)
+	if err != nil {
+		t.Fatalf("decode contact sheet PNG: %v", err)
+	}
+	if w := img.Bounds().Dx(); w < 100 {
+		t.Errorf("contact sheet looks empty (width %d)", w)
+	}
+}
+
+func TestCompareSkipContactSheet(t *testing.T) {
+	stderr, stdout, runDir, err := runCompareWithEnv(t, []string{
+		"x",
+		"--models", "mock",
+		"--size", "32x32",
+		"--seed", "1",
+		"--no-contact-sheet",
+	})
+	if err != nil {
+		t.Fatalf("runCompare: %v\nstderr: %s", err, stderr.String())
+	}
+	if got := strings.TrimSpace(stdout.String()); got != "" {
+		t.Errorf("expected no stdout output (no contact sheet), got %q", got)
+	}
+	if _, err := os.Stat(filepath.Join(runDir, "contact-sheet.png")); err == nil {
+		t.Errorf("contact-sheet.png should not exist with --no-contact-sheet")
+	}
+}
+
+func TestCompareRecordsBackendErrors(t *testing.T) {
+	// One real (mock) + one unknown. Unknown should fail but not abort the
+	// run — sidecar records both, contact sheet built from successes only.
+	stderr, _, runDir, err := runCompareWithEnv(t, []string{
+		"y",
+		"--models", "mock,this-instance-does-not-exist",
+		"--size", "32x32",
+	})
+	if err != nil {
+		t.Fatalf("runCompare: %v\nstderr: %s", err, stderr.String())
+	}
+	sidecar := filepath.Join(runDir, "compare.json")
+	data, _ := os.ReadFile(sidecar)
+	var body struct {
+		Successful int `json:"successful"`
+		Total      int `json:"total"`
+		Results    []struct {
+			Backend string `json:"backend"`
+			Error   string `json:"error"`
+		} `json:"results"`
+	}
+	if err := json.Unmarshal(data, &body); err != nil {
+		t.Fatalf("parse sidecar: %v", err)
+	}
+	if body.Total != 2 {
+		t.Errorf("expected 2 results, got %d", body.Total)
+	}
+	if body.Successful != 1 {
+		t.Errorf("expected 1 success, got %d", body.Successful)
+	}
+	var sawError bool
+	for _, r := range body.Results {
+		if r.Backend == "this-instance-does-not-exist" && r.Error != "" {
+			sawError = true
+		}
+	}
+	if !sawError {
+		t.Errorf("expected an error recorded for the unknown backend")
+	}
+}
+
+func TestCompareNoModelsFails(t *testing.T) {
+	_, _, _, err := runCompareWithEnv(t, []string{"x"})
+	if err == nil {
+		t.Fatal("expected error when --models is empty")
+	}
+	if !strings.Contains(err.Error(), "--models") {
+		t.Errorf("error should mention --models, got %v", err)
+	}
+}
+
+func TestCompareNoPromptFails(t *testing.T) {
+	_, _, _, err := runCompareWithEnv(t, []string{"--models", "mock"})
+	if err == nil {
+		t.Fatal("expected error when prompt is missing")
+	}
+	if !strings.Contains(err.Error(), "missing prompt") {
+		t.Errorf("error should mention missing prompt, got %v", err)
+	}
+}
diff --git a/cmd/imagen/main.go b/cmd/imagen/main.go
index 3913e44..b39151f 100644
--- a/cmd/imagen/main.go
+++ b/cmd/imagen/main.go
@@ -18,6 +18,8 @@ const helpText = `imagen — model-agnostic image generation
 
 Usage:
   imagen generate <prompt> [flags]    generate one image
+  imagen compare <prompt> --models a,b,c [flags]
+                                       run one prompt across N backends + contact sheet
   imagen worker [flags]               consume the imagen.jobs queue (daemon)
   imagen backends                     list registered backend types
   imagen config init                  print a sample imagen.yaml on stdout
@@ -46,6 +48,8 @@ func main() {
 	switch os.Args[1] {
 	case "generate":
 		err = runGenerate(ctx, args)
+	case "compare":
+		err = runCompare(ctx, args)
 	case "worker":
 		err = runWorker(ctx, args)
 	case "backends":
diff --git a/docs/backends.md b/docs/backends.md
new file mode 100644
index 0000000..cf1ea20
--- /dev/null
+++ b/docs/backends.md
@@ -0,0 +1,310 @@
+# ImaGen backends
+
+This document covers the local-ComfyUI backend plug-in story: how adapters
+are layered, how to add a new model without touching Go, and the per-model
+setup steps for the bundled templates.
+
+For the host-side ComfyUI install (mRock — venv, weights for the default
+FLUX.1-schnell, systemd, VRAM coexistence with Ollama, smoke test against
+the raw HTTP API), see [`setup-comfyui-mrock.md`](setup-comfyui-mrock.md).
+
+## Architecture: Path 1 — workflow-template adapter
+
+`imagen generate` and `imagen compare` dispatch through the `comfyui`
+adapter, which holds the HTTP plumbing (`/prompt`, `/history/{id}`, `/view`,
+`/system_stats`) and treats the workflow itself as data. Each backend
+instance in `imagen.yaml` picks a workflow JSON via the `workflow:` key.
+Adding a new model is yaml + JSON, never Go:
+
+```
+internal/backend/
+  comfyui.go              # one adapter, all ComfyUI models
+  workflow_template.go    # loader + token-substitution
+  workflows/
+    flux1-schnell.json    # bundled templates (embedded with //go:embed)
+    flux2-klein.json
+    sd35-medium.json
+```
+
+### Why Path 1 over per-family adapters (`comfyui-flux.go`, `comfyui-sd3.go`…)
+
+- **Workflow JSON is the natural exchange format**. ComfyUI users export
+  workflows from its GUI as JSON. Anything else means rebuilding the graph
+  by hand in Go for every new model.
+- **Adding a model is a config change, not a build change**. With Path 2,
+  every new family is a Go file, a new test file, a registry entry, a new
+  worker binary, a redeploy. Path 1 lets us land a new model with one yaml
+  block + one JSON file + one section in this doc.
+- **The HTTP plumbing is identical across families**. `/prompt`,
+  `/history`, `/view`, the retry policy, the "value not in list" hint, VRAM
+  reporting — none of it depends on the workflow shape. Path 2 would
+  duplicate that across files.
+- **Failure isolation stays clean**. The workflow loader fails at adapter
+  construction (`imagen backends` surfaces the error), the HTTP layer
+  fails at `Generate`, and ComfyUI's own validation surfaces missing-model
+  hints. Each layer's error message points at the right config knob.
+
+Path 2's argument was "each family owns its quirks (samplers, schedulers,
+dual-stage etc.)". That argument doesn't survive contact with the
+substitution-map design: per-family knobs are just key/value fields in the
+yaml block and `${shift}`/`${guidance}`/`${cfg}` placeholders in the
+template. No code duplication, no inheritance to debug.
+
+### Token substitution
+
+`workflow_template.SubstituteWorkflow` walks the parsed JSON and replaces
+every whole-value string of the form `"${key}"` with the typed value from
+the substitution map. Numbers stay numbers, strings stay strings — no
+round-tripping through `strings.Replace`.
+
+The substitution map is built per call from:
+
+1. **Request fields** (always present): `${prompt}`, `${negative}`,
+   `${width}`, `${height}`, `${seed}`, `${steps}`, `${sampler}`,
+   `${scheduler}`, `${cfg}`.
+2. **Every scalar field from the yaml block** (string / int / int64 /
+   float64 / bool), minus framework keys (`type`, `base_url`, `workflow`,
+   `default_*`). So `${vae}`, `${clip}`, `${clip_l}`, `${clip_t5}`,
+   `${dtype}`, `${shift}`, `${guidance}` all become substitutable just by
+   being in yaml.
+3. **Sensible defaults** for the common optional knobs above, so a
+   workflow that references `${dtype}` without the user setting one in
+   yaml still substitutes cleanly (`fp8_e4m3fn` for FLUX, `3.0` for SD3
+   shift, etc.). Extra defaults are ignored by workflows that don't
+   reference them.
+
+Partial matches (e.g. `"prefix ${prompt} suffix"`) are deliberately **not**
+substituted — the placeholder must be the entire value so we can preserve
+its JSON type. This prevents a prompt containing literal `${seed}` text
+from corrupting the workflow.
+
+Unknown placeholders (referenced in JSON but missing from the substitution
+map) error out before the workflow leaves the binary.
+
+### Back-compat
+
+The `workflow:` field defaults to `flux1-schnell` if omitted. Existing
+yaml blocks like the pre-#10 FLUX.1-schnell instance:
+
+```yaml
+flux-schnell-local:
+  type: comfyui
+  base_url: http://mrock:8188
+  model: flux1-schnell.safetensors
+```
+
+still work unchanged — they implicitly pick up the migrated
+`flux1-schnell.json` template, which keeps the same node IDs (6, 8, 9, 10,
+11, 12, 13, 27, 30, 31) as the historical hardcoded workflow.
+
+## Bundled workflows
+
+### FLUX.1-schnell — the back-compat default
+
+| Field | Default | Notes |
+|---|---|---|
+| `model` | `flux1-schnell.safetensors` | drop in `models/unet/` |
+| `vae` | `ae.safetensors` | `models/vae/` |
+| `clip_l` | `clip_l.safetensors` | `models/clip/` |
+| `clip_t5` | `t5xxl_fp8_e4m3fn.safetensors` | `models/clip/` |
+| `dtype` | `fp8_e4m3fn` | weight dtype for the UNet loader |
+| `default_steps` / `default_cfg` | 4 / 1.0 | schnell is distilled to ~4 steps |
+
+VRAM peak ~10–12 GB at 1024×1024. Install path:
+[`setup-comfyui-mrock.md`](setup-comfyui-mrock.md). Already shipping.
+
+### FLUX.2 [klein] 4B — direct upgrade
+
+Released by Black Forest Labs late 2025 / early 2026, BFL non-commercial
+license. The distilled 4B "klein" variant lands sub-second on the RTX
+4070 Ti SUPER and shares the new Qwen-based text encoder + a re-trained
+VAE with the larger family.
+
+```yaml
+flux2-klein-local:
+  type: comfyui
+  base_url: http://mrock:8188
+  workflow: flux2-klein
+  model: flux-2-klein-base-4b-fp8.safetensors    # models/unet/
+  vae: flux2-vae.safetensors                     # models/vae/
+  clip: qwen_3_4b.safetensors                    # models/text_encoders/
+  dtype: fp8_e4m3fn
+  default_steps: 4
+  default_cfg: 1.0
+  guidance: 4.0
+```
+
+**Model downloads** (on mRock, ungated mirrors when available):
+
+```bash
+cd ~/dev/comfyui/models
+curl -L -o unet/flux-2-klein-base-4b-fp8.safetensors \
+  https://huggingface.co/black-forest-labs/FLUX.2-klein/resolve/main/flux-2-klein-base-4b-fp8.safetensors
+curl -L -o vae/flux2-vae.safetensors \
+  https://huggingface.co/black-forest-labs/FLUX.2-klein/resolve/main/flux2-vae.safetensors
+mkdir -p text_encoders
+curl -L -o text_encoders/qwen_3_4b.safetensors \
+  https://huggingface.co/black-forest-labs/FLUX.2-klein/resolve/main/qwen_3_4b.safetensors
+```
+
+BFL's primary repo is gated; if `curl` returns 401, configure an HF token
+in `~/.cache/huggingface/token` or use one of the community mirrors
+(check the official model card for the current list). The filenames the
+template references match BFL's canonical names — rename downloads to
+match if a mirror uses different ones.
+
+VRAM peak: ~8.5 GB (4B fp8). With Ollama parked at ~8 GB this still fits;
+unlike FLUX.1-schnell, klein doesn't require stopping Ollama on mRock.
+
+### SD3.5-medium — single-checkpoint variant
+
+Stability AI's 2.5B mid-size model with bundled text encoders. The
+`incl_clips_t5xxlfp8scaled` variant ships clip_g + clip_l + t5xxl_fp8 all
+in one `.safetensors`, so the workflow uses `CheckpointLoaderSimple`
+instead of separate UNet/VAE/CLIP loaders.
+
+```yaml
+sd35-medium-local:
+  type: comfyui
+  base_url: http://mrock:8188
+  workflow: sd35-medium
+  model: sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors  # models/checkpoints/
+  default_steps: 28
+  default_sampler: dpmpp_2m
+  default_scheduler: sgm_uniform
+  default_cfg: 4.5
+  shift: 3.0
+```
+
+**Model download** (on mRock):
+
+```bash
+cd ~/dev/comfyui/models
+curl -L -o checkpoints/sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors \
+  https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/resolve/main/sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors
+```
+
+VRAM peak: ~9.9 GB at 1024×1024. Same envelope as FLUX.1-schnell — stop
+Ollama before generating, restart after.
+
+## Adding a new bundled workflow
+
+1. **Export from ComfyUI**: load the model in the ComfyUI GUI, build a
+   text-to-image workflow that produces what you want, "Save (API
+   Format)" — the file you get is the right shape.
+2. **Sprinkle placeholders**: open the JSON and replace per-call values
+   with `${name}` tokens. Whole-value substitution only:
+
+   ```json
+   "inputs": {
+     "text": "${prompt}",         // was "a cat sitting on a chair"
+     "seed": "${seed}",            // was 1234567
+     "steps": "${steps}",          // was 28
+     "cfg": "${cfg}",
+     "sampler_name": "${sampler}",
+     "scheduler": "${scheduler}",
+     "width": "${width}",
+     "height": "${height}"
+   }
+   ```
+
+   Use `${model}` for the checkpoint / unet filename and any per-template
+   knobs (`${vae}`, `${shift}`, `${guidance}`, `${clip}` …).
+3. **Drop it into `internal/backend/workflows/<name>.json`**. The
+   `//go:embed workflows/*.json` directive in `workflow_template.go`
+   picks it up at build time — no registry entry needed.
+4. **Add a yaml instance** in `internal/config/config.go`'s `Sample` block
+   for `imagen config init` (and `~/.config/imagen.yaml`) so users
+   discover the new backend.
+5. **Document the model files + HF download URLs** in this doc.
+6. **Smoke test**: `imagen generate "test" --backend <new-instance>
+   --size 1024x1024` should produce an image.
+
+Per-call overrides for sampler/scheduler/cfg go via `--steps`, `--seed`,
+and (programmatic) `backend.Request.BackendOpts["sampler"]` /
+`["scheduler"]` / `["cfg"]`. The compare harness forwards the
+constant-across-backends knobs verbatim.
+
+## Loading a workflow from disk (one-off)
+
+Pass an absolute filesystem path as `workflow:` and the adapter reads it
+from disk instead of the embedded FS. Handy for prototyping a new model
+before committing it:
+
+```yaml
+my-experimental:
+  type: comfyui
+  base_url: http://mrock:8188
+  workflow: /home/m/dev/comfyui/workflows/my-test.json
+  model: my-test-model.safetensors
+```
+
+The fallback chain is: filesystem path (if the string looks like a path
+or ends in `.json`), then bundled lookup by name, then bundled lookup
+with `.json` appended.
+
+## `imagen compare`: cross-backend evaluation
+
+```bash
+imagen compare "a wizard casting a spell" \
+  --models flux-schnell-local,flux2-klein-local,sd35-medium-local \
+  --size 1024x1024 \
+  --output ~/Pictures/imagen/compare
+```
+
+Per run, `compare`:
+
+- creates `<output>/<YYYYMMDD-HHMMSS>-<prompt-slug>/`
+- dispatches each named backend sequentially (mRock has one GPU; parallel
+  would OOM) — one backend's failure doesn't abort the run
+- writes per-backend PNGs as `<prompt-slug>--<backend-slug>.png`
+- writes `compare.json` listing every attempt (success + failure) with
+  per-model `seed`, `latency_ms`, `model`, `vram_used_mib`, full
+  `metadata` map, and the error string for any failure
+- composites a `contact-sheet.png` with the prompt as header and each
+  cell labelled `<backend>` / `<latency>ms · seed <n>`
+
+Flags mirror `generate`: `--seed`, `--steps`, `--style`, `--negative`,
+`--size` are shared across all backends. `--no-contact-sheet` skips the
+composite when only the per-image PNGs and sidecar matter (e.g. for a
+worker script that builds its own diff view).
+
+## Diagnostics
+
+`imagen backends` shows every instance with its registration state. For
+local ComfyUI, the status is currently just `registered` (we don't probe
+the upstream HTTP endpoint at startup — the boot-helper hint kicks in on
+first generation if mRock is asleep).
+
+Per-backend errors emit at most three kinds:
+
+1. **Adapter construction failure** (e.g. workflow JSON not found,
+   missing required yaml field). Caught at `buildBackend` time:
+   `imagen: backend "<name>": <err>`.
+2. **HTTP / runtime failure during Generate**. Wrapped with the boot
+   helper for `connection refused`/`no such host`/timeouts pointing at
+   `boot-whitetower mrock` so a sleeping mRock has an obvious next step.
+3. **ComfyUI workflow-validation failure** (200-with-node_errors or 400).
+   Surfaces with a model-not-found hint (matching `value_not_in_list` +
+   `unet_name`/`ckpt_name`) when applicable, pointing back at this doc.
+
+## Worker daemon notes
+
+`imagen worker` (the `imagen.jobs` queue consumer) uses the same adapter
++ workflow lookup as the synchronous CLI — flexsiebels' `/imagine` UI
+INSERTs a `backend = <instance>` row, the worker claims it, and the
+underlying ComfyUI HTTP calls are identical to what `generate` makes. No
+worker-specific changes are required when a new backend lands; the
+config + workflow are the only state that has to be present on the
+worker host.
+
+After merging a new template or yaml block:
+
+```bash
+# On the worker host (mRiver today):
+systemctl --user restart imagen-worker
+```
+
+The daemon-rebuild trap from issue #9 still applies: if you build the
+imagen binary on the dev machine and `scp` it over, restart the unit so
+systemd picks up the new ELF.
diff --git a/docs/usage.md b/docs/usage.md
index 75585fc..a69c93c 100644
--- a/docs/usage.md
+++ b/docs/usage.md
@@ -4,14 +4,21 @@
 
 ```
 imagen generate <prompt> [flags]    generate one image
+imagen compare <prompt> --models a,b,c [flags]
+                                    run one prompt across N backends + contact sheet
+imagen worker [flags]               consume the imagen.jobs queue (daemon)
 imagen backends                     list configured + registered backends
 imagen config init                  print a sample imagen.yaml on stdout
 imagen config validate              parse + validate the active config
 imagen config path                  print the resolved config path
 imagen serve [--addr :8080]         (stub) start the HTTP server
+imagen usage [--since DATE]         show cost-tracking rows
 imagen version                      print version
 ```
 
+For the per-backend setup (FLUX.1, FLUX.2 [klein], SD3.5 medium, …) and
+the architecture rationale, see [`backends.md`](backends.md).
+
 ## `generate` flags
 
 | Flag           | Default                      | Notes                                                       |
diff --git a/go.mod b/go.mod
index 5cd50ab..1d9db0b 100644
--- a/go.mod
+++ b/go.mod
@@ -4,6 +4,7 @@ go 1.25.0
 
 require (
 	github.com/jackc/pgx/v5 v5.9.2
+	golang.org/x/image v0.40.0
 	gopkg.in/yaml.v3 v3.0.1
 )
 
@@ -12,5 +13,5 @@ require (
 	github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761 // indirect
 	github.com/kr/text v0.2.0 // indirect
 	github.com/rogpeppe/go-internal v1.14.1 // indirect
-	golang.org/x/text v0.29.0 // indirect
+	golang.org/x/text v0.37.0 // indirect
 )
diff --git a/go.sum b/go.sum
index 7af349d..70d691b 100644
--- a/go.sum
+++ b/go.sum
@@ -23,10 +23,12 @@ github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UV
 github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
 github.com/stretchr/testify v1.11.1 h1:7s2iGBzp5EwR7/aIZr8ao5+dra3wiQyKjjFuvgVKu7U=
 github.com/stretchr/testify v1.11.1/go.mod h1:wZwfW3scLgRK+23gO65QZefKpKQRnfz6sD981Nm4B6U=
-golang.org/x/sync v0.17.0 h1:l60nONMj9l5drqw6jlhIELNv9I0A4OFgRsG9k2oT9Ug=
-golang.org/x/sync v0.17.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=
-golang.org/x/text v0.29.0 h1:1neNs90w9YzJ9BocxfsQNHKuAT4pkghyXc4nhZ6sJvk=
-golang.org/x/text v0.29.0/go.mod h1:7MhJOA9CD2qZyOKYazxdYMF85OwPdEr9jTtBpO7ydH4=
+golang.org/x/image v0.40.0 h1:Tw4GyDXMo+daZN1znreBRC3VayR1aLFUyUEOLUdW1a8=
+golang.org/x/image v0.40.0/go.mod h1:uIc348UZMSvS5Z65CVZ7iDPaNobNFEPeJ4kbqTOszmA=
+golang.org/x/sync v0.20.0 h1:e0PTpb7pjO8GAtTs2dQ6jYa5BWYlMuX047Dco/pItO4=
+golang.org/x/sync v0.20.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/text v0.37.0 h1:Cqjiwd9eSg8e0QAkyCaQTNHFIIzWtidPahFWR83rTrc=
+golang.org/x/text v0.37.0/go.mod h1:a5sjxXGs9hsn/AJVwuElvCAo9v8QYLzvavO5z2PiM38=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=
 gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=
diff --git a/internal/backend/comfyui.go b/internal/backend/comfyui.go
index 442e1e2..fdfd50e 100644
--- a/internal/backend/comfyui.go
+++ b/internal/backend/comfyui.go
@@ -20,24 +20,29 @@ import (
 const ComfyType = "comfyui"
 
 // Comfy is the ComfyUI adapter. It speaks the public `/prompt` + `/history`
-// + `/view` HTTP API and submits a fixed FLUX.1 schnell workflow built from
-// the values in Request.
+// + `/view` HTTP API and submits a workflow built by substituting Request
+// values into a JSON template (bundled under internal/backend/workflows/ or
+// loaded from a filesystem path).
 //
 // Concurrency: a single Comfy is safe to share across goroutines as long as
 // the underlying http.Client is. Generate does not hold long-lived state.
 type Comfy struct {
 	instance string
 
-	base    string
-	model   string
-	vae     string
-	clipL   string
-	clipT5  string
-	dtype   string
+	base     string
+	workflow string
+
+	// rawCfg keeps the original yaml block (minus framework keys) so we can
+	// expose every user-defined string/number as a workflow substitution
+	// without enumerating each per-model knob in Go. Empty values still get
+	// a substitution entry so a template can reference ${negative} when the
+	// request didn't pass one.
+	rawCfg map[string]any
 
 	defaultSteps     int
 	defaultSampler   string
 	defaultScheduler string
+	defaultCFG       float64
 
 	httpClient   *http.Client
 	pollInterval time.Duration
@@ -49,12 +54,20 @@ type Comfy struct {
 }
 
 // NewComfy is the registry constructor. cfg is the adapter's slice of
-// imagen.yaml. Required keys: base_url, model. The rest have sensible FLUX
-// schnell defaults.
+// imagen.yaml.
+//
+// Required keys: base_url, model.
+// Optional keys: workflow (defaults to "flux1-schnell" for back-compat with
+// existing configs), default_steps, default_sampler, default_scheduler,
+// default_cfg, plus any template-specific knobs (vae, clip, clip_l,
+// clip_t5, dtype, shift, guidance, …) the chosen workflow references.
 func NewComfy(name string, cfg map[string]any) (Backend, error) {
 	if name == "" {
 		return nil, fmt.Errorf("comfyui: empty instance name")
 	}
+	if cfg == nil {
+		cfg = map[string]any{}
+	}
 	base := strings.TrimRight(getString(cfg, "base_url", ""), "/")
 	if base == "" {
 		return nil, fmt.Errorf("comfyui[%s]: base_url is required", name)
@@ -67,23 +80,27 @@ func NewComfy(name string, cfg map[string]any) (Backend, error) {
 		return nil, fmt.Errorf("comfyui[%s]: model is required", name)
 	}
 
+	workflow := getString(cfg, "workflow", "flux1-schnell")
+	// Fail fast on a bad workflow ref so users see the error at startup,
+	// not on first /prompt submission.
+	if _, err := LoadWorkflowTemplate(workflow); err != nil {
+		return nil, fmt.Errorf("comfyui[%s]: %w", name, err)
+	}
+
 	c := &Comfy{
 		instance: name,
 		base:     base,
-		model:    model,
-
-		vae:    getString(cfg, "vae", "ae.safetensors"),
-		clipL:  getString(cfg, "clip_l", "clip_l.safetensors"),
-		clipT5: getString(cfg, "clip_t5", "t5xxl_fp8_e4m3fn.safetensors"),
-		dtype:  getString(cfg, "weight_dtype", "fp8_e4m3fn"),
+		workflow: workflow,
+		rawCfg:   cfg,
 
 		defaultSteps:     getInt(cfg, "default_steps", 4),
 		defaultSampler:   getString(cfg, "default_sampler", "euler"),
 		defaultScheduler: getString(cfg, "default_scheduler", "simple"),
+		defaultCFG:       getFloat(cfg, "default_cfg", 1.0),
 
 		httpClient:   &http.Client{Timeout: 60 * time.Second},
 		pollInterval: 250 * time.Millisecond,
-		pollTimeout:  120 * time.Second,
+		pollTimeout:  300 * time.Second,
 
 		randSeed:   cryptoSeed,
 		clientIDFn: randClientID,
@@ -103,19 +120,26 @@ func (c *Comfy) Generate(ctx context.Context, req Request) (*Result, error) {
 
 	sampler := c.defaultSampler
 	scheduler := c.defaultScheduler
+	cfg := c.defaultCFG
 	if v, ok := req.BackendOpts["sampler"].(string); ok && v != "" {
 		sampler = v
 	}
 	if v, ok := req.BackendOpts["scheduler"].(string); ok && v != "" {
 		scheduler = v
 	}
+	if v, ok := req.BackendOpts["cfg"].(float64); ok && v > 0 {
+		cfg = v
+	}
 
 	seed := req.Seed
 	if seed == 0 {
 		seed = c.randSeed()
 	}
 
-	workflow := c.buildWorkflow(req.Prompt, req.NegativePrompt, width, height, seed, steps, sampler, scheduler)
+	workflow, err := c.buildWorkflow(req.Prompt, req.NegativePrompt, width, height, seed, steps, sampler, scheduler, cfg)
+	if err != nil {
+		return nil, fmt.Errorf("comfyui[%s]: build workflow: %w", c.instance, err)
+	}
 	clientID := c.clientIDFn()
 
 	start := time.Now()
@@ -133,14 +157,17 @@ func (c *Comfy) Generate(ctx context.Context, req Request) (*Result, error) {
 	}
 	latencyMs := time.Since(start).Milliseconds()
 
+	model := getString(c.rawCfg, "model", "")
 	meta := map[string]any{
 		"backend":      c.instance,
 		"backend_type": ComfyType,
-		"model":        c.model,
+		"workflow":     c.workflow,
+		"model":        model,
 		"seed":         seed,
 		"steps":        steps,
 		"sampler":      sampler,
 		"scheduler":    scheduler,
+		"cfg":          cfg,
 		"width":        width,
 		"height":       height,
 		"latency_ms":   latencyMs,
@@ -173,6 +200,7 @@ func (c *Comfy) submitPrompt(ctx context.Context, workflow map[string]any, clien
 		return "", fmt.Errorf("comfyui: marshal workflow: %w", err)
 	}
 
+	model := getString(c.rawCfg, "model", "")
 	var lastErr error
 	for attempt := range 2 {
 		if attempt > 0 {
@@ -196,7 +224,7 @@ func (c *Comfy) submitPrompt(ctx context.Context, workflow map[string]any, clien
 		_ = resp.Body.Close()
 		switch {
 		case resp.StatusCode >= 200 && resp.StatusCode < 300:
-			return parsePromptID(respBody, c.model)
+			return parsePromptID(respBody, model)
 		case resp.StatusCode >= 500:
 			lastErr = fmt.Errorf("comfyui /prompt %d: %s", resp.StatusCode, snip(respBody))
 			continue
@@ -333,98 +361,74 @@ func (c *Comfy) connError(err error) error {
 // workflow-validation failures and put the diagnostics in node_errors; older
 // builds use 200 + node_errors. This handles the 4xx flavour.
 func (c *Comfy) classifyBadRequest(status int, body []byte) error {
-	if hint, ok := missingModelHint(body, c.model); ok {
-		return fmt.Errorf("comfyui /prompt %d: %s — see docs/setup-comfyui-mrock.md", status, hint)
+	model := getString(c.rawCfg, "model", "")
+	if hint, ok := missingModelHint(body, model); ok {
+		return fmt.Errorf("comfyui /prompt %d: %s — see docs/backends.md", status, hint)
 	}
 	return fmt.Errorf("comfyui /prompt %d: %s", status, snip(body))
 }
 
-// buildWorkflow assembles the canonical FLUX.1 schnell ComfyUI workflow,
-// node-IDs matching the upstream "flux-schnell" template so anyone debugging
-// in the ComfyUI UI sees a familiar shape.
-func (c *Comfy) buildWorkflow(prompt, negative string, w, h int, seed int64, steps int, sampler, scheduler string) map[string]any {
-	return map[string]any{
-		"6": map[string]any{
-			"class_type": "CLIPTextEncode",
-			"inputs": map[string]any{
-				"text": prompt,
-				"clip": []any{"11", 0},
-			},
-		},
-		"8": map[string]any{
-			"class_type": "VAEDecode",
-			"inputs": map[string]any{
-				"samples": []any{"31", 0},
-				"vae":     []any{"10", 0},
-			},
-		},
-		"9": map[string]any{
-			"class_type": "SaveImage",
-			"inputs": map[string]any{
-				"filename_prefix": "imagen",
-				"images":          []any{"8", 0},
-			},
-		},
-		"10": map[string]any{
-			"class_type": "VAELoader",
-			"inputs":     map[string]any{"vae_name": c.vae},
-		},
-		"11": map[string]any{
-			"class_type": "DualCLIPLoader",
-			"inputs": map[string]any{
-				"clip_name1": c.clipT5,
-				"clip_name2": c.clipL,
-				"type":       "flux",
-			},
-		},
-		"12": map[string]any{
-			"class_type": "UNETLoader",
-			"inputs": map[string]any{
-				"unet_name":    c.model,
-				"weight_dtype": c.dtype,
-			},
-		},
-		"13": map[string]any{
-			"class_type": "CLIPTextEncode",
-			"inputs": map[string]any{
-				"text": negative,
-				"clip": []any{"11", 0},
-			},
-		},
-		"27": map[string]any{
-			"class_type": "EmptySD3LatentImage",
-			"inputs": map[string]any{
-				"width":      w,
-				"height":     h,
-				"batch_size": 1,
-			},
-		},
-		"30": map[string]any{
-			"class_type": "ModelSamplingFlux",
-			"inputs": map[string]any{
-				"model":      []any{"12", 0},
-				"max_shift":  1.15,
-				"base_shift": 0.5,
-				"width":      w,
-				"height":     h,
-			},
-		},
-		"31": map[string]any{
-			"class_type": "KSampler",
-			"inputs": map[string]any{
-				"model":        []any{"30", 0},
-				"seed":         seed,
-				"steps":        steps,
-				"cfg":          1.0,
-				"sampler_name": sampler,
-				"scheduler":    scheduler,
-				"denoise":      1.0,
-				"positive":     []any{"6", 0},
-				"negative":     []any{"13", 0},
-				"latent_image": []any{"27", 0},
-			},
-		},
+// buildWorkflow loads the configured workflow template and substitutes the
+// per-call placeholders (prompt, seed, sampler, …) plus any string/number
+// fields the user defined in the yaml block. The set of placeholder keys
+// that aren't in `subs` produces an error from SubstituteWorkflow.
+func (c *Comfy) buildWorkflow(prompt, negative string, w, h int, seed int64, steps int, sampler, scheduler string, cfg float64) (map[string]any, error) {
+	wf, err := LoadWorkflowTemplate(c.workflow)
+	if err != nil {
+		return nil, err
 	}
+	subs := map[string]any{
+		"prompt":    prompt,
+		"negative":  negative,
+		"width":     w,
+		"height":    h,
+		"seed":      seed,
+		"steps":     steps,
+		"sampler":   sampler,
+		"scheduler": scheduler,
+		"cfg":       cfg,
+	}
+	// Surface every scalar field from the yaml block so per-template knobs
+	// (vae, clip, clip_l, clip_t5, dtype, shift, guidance, …) work without
+	// adapter-code changes. Framework keys are excluded.
+	for k, v := range c.rawCfg {
+		switch k {
+		case "type", "base_url", "workflow",
+			"default_steps", "default_sampler", "default_scheduler", "default_cfg":
+			continue
+		}
+		if _, alreadySet := subs[k]; alreadySet {
+			// A per-call var (e.g. ${prompt}) beats anything yaml put under
+			// the same key — yaml can't shadow request-derived values.
+			continue
+		}
+		switch v := v.(type) {
+		case string, int, int64, float64, bool:
+			subs[k] = v
+		}
+	}
+	// Provide sensible defaults for common optional knobs so a workflow that
+	// references one of these doesn't fail substitution when the user
+	// didn't override it in yaml. Extra keys are ignored if the workflow
+	// doesn't reference them, so it's safe to always set the lot.
+	defaults := map[string]any{
+		"vae":      "ae.safetensors",
+		"clip_l":   "clip_l.safetensors",
+		"clip_t5":  "t5xxl_fp8_e4m3fn.safetensors",
+		"clip":     "qwen_3_4b.safetensors",
+		"dtype":    "fp8_e4m3fn",
+		"guidance": 4.0,
+		"shift":    3.0,
+	}
+	for k, v := range defaults {
+		if _, ok := subs[k]; !ok {
+			subs[k] = v
+		}
+	}
+	if _, err := SubstituteWorkflow(wf, subs); err != nil {
+		return nil, err
+	}
+	return wf, nil
 }
 
 // parsePromptID handles the 2xx /prompt response. ComfyUI sometimes 200s a
@@ -432,8 +436,8 @@ func (c *Comfy) buildWorkflow(prompt, negative string, w, h int, seed int64, ste
 // turns that into the same user-facing error as a 4xx with the same body.
 func parsePromptID(body []byte, model string) (string, error) {
 	var resp struct {
-		PromptID   string         `json:"prompt_id"`
-		NodeErrors map[string]any `json:"node_errors"`
+		PromptID   string          `json:"prompt_id"`
+		NodeErrors map[string]any  `json:"node_errors"`
 		Error      json.RawMessage `json:"error"`
 	}
 	if err := json.Unmarshal(body, &resp); err != nil {
@@ -441,7 +445,7 @@ func parsePromptID(body []byte, model string) (string, error) {
 	}
 	if len(resp.NodeErrors) > 0 || len(resp.Error) > 0 {
 		if hint, ok := missingModelHint(body, model); ok {
-			return "", fmt.Errorf("comfyui /prompt: %s — see docs/setup-comfyui-mrock.md", hint)
+			return "", fmt.Errorf("comfyui /prompt: %s — see docs/backends.md", hint)
 		}
 		return "", fmt.Errorf("comfyui /prompt rejected workflow: %s", snip(body))
 	}
@@ -489,15 +493,21 @@ func parseHistory(body []byte, promptID string) (string, bool, error) {
 }
 
 // missingModelHint returns a user-actionable message when the response body
-// indicates the configured unet model isn't loaded on the server. ComfyUI
-// uses both the human-readable "Value not in list" message and the enum
-// "value_not_in_list" type — match either.
+// indicates the configured unet/checkpoint model isn't loaded on the server.
+// ComfyUI uses both the human-readable "Value not in list" message and the
+// enum "value_not_in_list" type — match either.
 func missingModelHint(body []byte, model string) (string, bool) {
 	s := string(body)
 	hasMarker := strings.Contains(s, "Value not in list") || strings.Contains(s, "value_not_in_list")
-	if hasMarker && strings.Contains(s, "unet_name") {
+	if !hasMarker {
+		return "", false
+	}
+	if strings.Contains(s, "unet_name") {
 		return fmt.Sprintf("model %q not present in the ComfyUI server's models/unet/", model), true
 	}
+	if strings.Contains(s, "ckpt_name") {
+		return fmt.Sprintf("checkpoint %q not present in the ComfyUI server's models/checkpoints/", model), true
+	}
 	return "", false
 }
 
@@ -536,6 +546,22 @@ func getInt(m map[string]any, k string, def int) int {
 	return def
 }
 
+func getFloat(m map[string]any, k string, def float64) float64 {
+	if v, ok := m[k]; ok {
+		switch n := v.(type) {
+		case float64:
+			return n
+		case float32:
+			return float64(n)
+		case int:
+			return float64(n)
+		case int64:
+			return float64(n)
+		}
+	}
+	return def
+}
+
 func orDefaultInt(v, def int) int {
 	if v == 0 {
 		return def
diff --git a/internal/backend/comfyui_test.go b/internal/backend/comfyui_test.go
index 778961d..0d66879 100644
--- a/internal/backend/comfyui_test.go
+++ b/internal/backend/comfyui_test.go
@@ -312,7 +312,7 @@ func TestComfyMissingModelHintsAtSetupDoc(t *testing.T) {
 		t.Fatal("expected error")
 	}
 	msg := err.Error()
-	if !strings.Contains(msg, "docs/setup-comfyui-mrock.md") {
+	if !strings.Contains(msg, "docs/backends.md") {
 		t.Errorf("error should point at the setup doc, got %v", err)
 	}
 	if !strings.Contains(msg, "flux1-schnell.safetensors") {
@@ -331,7 +331,7 @@ func TestComfyMissingModelOn200WithNodeErrors(t *testing.T) {
 	if err == nil {
 		t.Fatal("expected error for node_errors on 200")
 	}
-	if !strings.Contains(err.Error(), "docs/setup-comfyui-mrock.md") {
+	if !strings.Contains(err.Error(), "docs/backends.md") {
 		t.Errorf("error should point at the setup doc, got %v", err)
 	}
 }
diff --git a/internal/backend/workflow_template.go b/internal/backend/workflow_template.go
new file mode 100644
index 0000000..806e655
--- /dev/null
+++ b/internal/backend/workflow_template.go
@@ -0,0 +1,156 @@
+package backend
+
+import (
+	"embed"
+	"encoding/json"
+	"fmt"
+	"io/fs"
+	"maps"
+	"os"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"strings"
+)
+
+//go:embed workflows/*.json
+var bundledWorkflows embed.FS
+
+// placeholderRE matches a single-token placeholder like "${prompt}" — the
+// whole string value must be the placeholder, leading/trailing whitespace
+// allowed. This lets us preserve types (a numeric substitution becomes a
+// JSON number, not a stringified one) instead of round-tripping through
+// strings.Replace which would force everything into a string.
+var placeholderRE = regexp.MustCompile(`^\s*\$\{([a-zA-Z][a-zA-Z0-9_]*)\}\s*$`)
+
+// LoadWorkflowTemplate returns the parsed JSON for a workflow template.
+// `name` is resolved in this order:
+//
+//  1. exact filesystem path that exists on disk (absolute or relative);
+//  2. one of the bundled templates under internal/backend/workflows/
+//     (with or without the .json suffix).
+//
+// The returned map is a fresh deep copy of the template; callers can mutate
+// it freely.
+func LoadWorkflowTemplate(name string) (map[string]any, error) {
+	if name == "" {
+		return nil, fmt.Errorf("workflow template name is empty")
+	}
+	raw, err := readWorkflowBytes(name)
+	if err != nil {
+		return nil, err
+	}
+	var wf map[string]any
+	if err := json.Unmarshal(raw, &wf); err != nil {
+		return nil, fmt.Errorf("workflow %s: parse: %w", name, err)
+	}
+	return wf, nil
+}
+
+// BundledWorkflowNames returns the names of templates compiled into the
+// binary, sorted. Each name is the basename without the .json suffix.
+func BundledWorkflowNames() []string {
+	entries, err := fs.ReadDir(bundledWorkflows, "workflows")
+	if err != nil {
+		return nil
+	}
+	out := make([]string, 0, len(entries))
+	for _, e := range entries {
+		n := e.Name()
+		if !strings.HasSuffix(n, ".json") {
+			continue
+		}
+		out = append(out, strings.TrimSuffix(n, ".json"))
+	}
+	sort.Strings(out)
+	return out
+}
+
+func readWorkflowBytes(name string) ([]byte, error) {
+	// Filesystem path wins if it points at a real file. Lets a user override
+	// a bundled template by passing an absolute path in yaml.
+	if strings.ContainsRune(name, os.PathSeparator) || strings.HasSuffix(name, ".json") {
+		if b, err := os.ReadFile(name); err == nil {
+			return b, nil
+		} else if !os.IsNotExist(err) {
+			return nil, fmt.Errorf("workflow %s: %w", name, err)
+		}
+	}
+	// Bundled lookup. Try the literal name as a file inside workflows/, then
+	// with the .json suffix appended.
+	candidates := []string{
+		filepath.Join("workflows", name),
+		filepath.Join("workflows", name+".json"),
+	}
+	for _, c := range candidates {
+		if b, err := bundledWorkflows.ReadFile(c); err == nil {
+			return b, nil
+		}
+	}
+	return nil, fmt.Errorf("workflow %q not found (bundled templates: %v)", name, BundledWorkflowNames())
+}
+
+// SubstituteWorkflow walks wf and replaces every "${key}" string with the
+// matching value from subs, preserving JSON types. Returns the set of
+// placeholder keys it actually touched, so the caller can detect missing
+// substitutions even when a key is defined in subs but never referenced in
+// the workflow (typical when a yaml block sets a knob a different template
+// would consume).
+//
+// Unknown placeholders (referenced in the workflow but absent from subs)
+// produce an error so we never submit a workflow with raw "${foo}" tokens.
+func SubstituteWorkflow(wf map[string]any, subs map[string]any) (used map[string]struct{}, err error) {
+	used = make(map[string]struct{})
+	walked, err := substituteValue(wf, subs, used)
+	if err != nil {
+		return nil, err
+	}
+	// substituteValue returns the replacement for the top-level value, which
+	// should still be the same map (just with mutated children).
+	if m, ok := walked.(map[string]any); ok {
+		// Copy back into wf so the caller's reference reflects the result.
+		for k := range wf {
+			delete(wf, k)
+		}
+		maps.Copy(wf, m)
+	}
+	return used, nil
+}
+
+func substituteValue(v any, subs map[string]any, used map[string]struct{}) (any, error) {
+	switch x := v.(type) {
+	case map[string]any:
+		out := make(map[string]any, len(x))
+		for k, child := range x {
+			replaced, err := substituteValue(child, subs, used)
+			if err != nil {
+				return nil, err
+			}
+			out[k] = replaced
+		}
+		return out, nil
+	case []any:
+		out := make([]any, len(x))
+		for i, child := range x {
+			replaced, err := substituteValue(child, subs, used)
+			if err != nil {
+				return nil, err
+			}
+			out[i] = replaced
+		}
+		return out, nil
+	case string:
+		if m := placeholderRE.FindStringSubmatch(x); m != nil {
+			key := m[1]
+			val, ok := subs[key]
+			if !ok {
+				return nil, fmt.Errorf("workflow placeholder ${%s} has no substitution", key)
+			}
+			used[key] = struct{}{}
+			return val, nil
+		}
+		return x, nil
+	default:
+		return v, nil
+	}
+}
diff --git a/internal/backend/workflow_template_test.go b/internal/backend/workflow_template_test.go
new file mode 100644
index 0000000..dbd8ffe
--- /dev/null
+++ b/internal/backend/workflow_template_test.go
@@ -0,0 +1,153 @@
+package backend
+
+import (
+	"os"
+	"path/filepath"
+	"slices"
+	"strings"
+	"testing"
+)
+
+func TestBundledWorkflowsParseable(t *testing.T) {
+	names := BundledWorkflowNames()
+	if len(names) == 0 {
+		t.Fatal("expected at least one bundled workflow")
+	}
+	mustHave := []string{"flux1-schnell", "flux2-klein", "sd35-medium"}
+	for _, want := range mustHave {
+		if !slices.Contains(names, want) {
+			t.Errorf("bundled workflows missing %q (have: %v)", want, names)
+		}
+	}
+	// Every bundled template must parse and contain at least one node.
+	for _, n := range names {
+		wf, err := LoadWorkflowTemplate(n)
+		if err != nil {
+			t.Errorf("LoadWorkflowTemplate(%q): %v", n, err)
+			continue
+		}
+		if len(wf) == 0 {
+			t.Errorf("workflow %q has zero nodes", n)
+		}
+	}
+}
+
+func TestLoadWorkflowFromFilesystem(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "custom.json")
+	body := `{"1":{"class_type":"X","inputs":{"v":"${prompt}"}}}`
+	if err := os.WriteFile(path, []byte(body), 0o644); err != nil {
+		t.Fatalf("write tmp workflow: %v", err)
+	}
+	wf, err := LoadWorkflowTemplate(path)
+	if err != nil {
+		t.Fatalf("load from path: %v", err)
+	}
+	if _, ok := wf["1"]; !ok {
+		t.Errorf("custom workflow missing node 1")
+	}
+}
+
+func TestLoadWorkflowUnknownNameErrors(t *testing.T) {
+	_, err := LoadWorkflowTemplate("definitely-not-a-real-workflow")
+	if err == nil {
+		t.Fatal("expected error for unknown workflow name")
+	}
+	if !strings.Contains(err.Error(), "not found") {
+		t.Errorf("error should say not found, got %v", err)
+	}
+}
+
+func TestSubstituteWorkflowPreservesTypes(t *testing.T) {
+	wf := map[string]any{
+		"31": map[string]any{
+			"class_type": "KSampler",
+			"inputs": map[string]any{
+				"seed":  "${seed}",
+				"steps": "${steps}",
+				"text":  "${prompt}",
+				"cfg":   "${cfg}",
+			},
+		},
+	}
+	subs := map[string]any{
+		"seed":   int64(42),
+		"steps":  11,
+		"prompt": "a cat",
+		"cfg":    4.5,
+	}
+	used, err := SubstituteWorkflow(wf, subs)
+	if err != nil {
+		t.Fatalf("Substitute: %v", err)
+	}
+	if len(used) != 4 {
+		t.Errorf("used = %v, want all four", used)
+	}
+	inputs := wf["31"].(map[string]any)["inputs"].(map[string]any)
+	if seed, ok := inputs["seed"].(int64); !ok || seed != 42 {
+		t.Errorf("seed = %T %v, want int64 42", inputs["seed"], inputs["seed"])
+	}
+	if steps, ok := inputs["steps"].(int); !ok || steps != 11 {
+		t.Errorf("steps = %T %v, want int 11", inputs["steps"], inputs["steps"])
+	}
+	if text, ok := inputs["text"].(string); !ok || text != "a cat" {
+		t.Errorf("text = %T %v, want string", inputs["text"], inputs["text"])
+	}
+	if cfg, ok := inputs["cfg"].(float64); !ok || cfg != 4.5 {
+		t.Errorf("cfg = %T %v, want float64 4.5", inputs["cfg"], inputs["cfg"])
+	}
+}
+
+func TestSubstituteWorkflowMissingPlaceholderErrors(t *testing.T) {
+	wf := map[string]any{
+		"1": map[string]any{"inputs": map[string]any{"v": "${missing}"}},
+	}
+	_, err := SubstituteWorkflow(wf, map[string]any{})
+	if err == nil {
+		t.Fatal("expected error for missing placeholder")
+	}
+	if !strings.Contains(err.Error(), "${missing}") {
+		t.Errorf("error should name the placeholder, got %v", err)
+	}
+}
+
+func TestSubstituteWorkflowOnlyWholeTokens(t *testing.T) {
+	// Partial-match strings ("prefix ${prompt} suffix") are NOT substituted —
+	// the placeholder must be the whole value so we can preserve types.
+	wf := map[string]any{
+		"1": map[string]any{"inputs": map[string]any{
+			"keep_string":  "stuff with ${prompt} inside",
+			"replace_full": "${prompt}",
+		}},
+	}
+	used, err := SubstituteWorkflow(wf, map[string]any{"prompt": "x"})
+	if err != nil {
+		t.Fatalf("Substitute: %v", err)
+	}
+	inputs := wf["1"].(map[string]any)["inputs"].(map[string]any)
+	if inputs["keep_string"].(string) != "stuff with ${prompt} inside" {
+		t.Errorf("partial match should be left alone, got %q", inputs["keep_string"])
+	}
+	if inputs["replace_full"].(string) != "x" {
+		t.Errorf("full-value match should substitute, got %q", inputs["replace_full"])
+	}
+	if _, ok := used["prompt"]; !ok {
+		t.Errorf("used should track keys that fired")
+	}
+}
+
+func TestFlux1SchnellTemplateMatchesLegacyShape(t *testing.T) {
+	// Regression guard against the historical hardcoded workflow: every
+	// node ID the old Comfy.buildWorkflow used must still exist in the
+	// migrated template.
+	wf, err := LoadWorkflowTemplate("flux1-schnell")
+	if err != nil {
+		t.Fatalf("load flux1-schnell: %v", err)
+	}
+	legacyNodes := []string{"6", "8", "9", "10", "11", "12", "13", "27", "30", "31"}
+	for _, id := range legacyNodes {
+		if _, ok := wf[id]; !ok {
+			t.Errorf("flux1-schnell template missing node %q (legacy parity)", id)
+		}
+	}
+}
diff --git a/internal/backend/workflows/flux1-schnell.json b/internal/backend/workflows/flux1-schnell.json
new file mode 100644
index 0000000..a137d4e
--- /dev/null
+++ b/internal/backend/workflows/flux1-schnell.json
@@ -0,0 +1,84 @@
+{
+  "6": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "text": "${prompt}",
+      "clip": ["11", 0]
+    }
+  },
+  "8": {
+    "class_type": "VAEDecode",
+    "inputs": {
+      "samples": ["31", 0],
+      "vae": ["10", 0]
+    }
+  },
+  "9": {
+    "class_type": "SaveImage",
+    "inputs": {
+      "filename_prefix": "imagen",
+      "images": ["8", 0]
+    }
+  },
+  "10": {
+    "class_type": "VAELoader",
+    "inputs": {
+      "vae_name": "${vae}"
+    }
+  },
+  "11": {
+    "class_type": "DualCLIPLoader",
+    "inputs": {
+      "clip_name1": "${clip_t5}",
+      "clip_name2": "${clip_l}",
+      "type": "flux"
+    }
+  },
+  "12": {
+    "class_type": "UNETLoader",
+    "inputs": {
+      "unet_name": "${model}",
+      "weight_dtype": "${dtype}"
+    }
+  },
+  "13": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "text": "${negative}",
+      "clip": ["11", 0]
+    }
+  },
+  "27": {
+    "class_type": "EmptySD3LatentImage",
+    "inputs": {
+      "width": "${width}",
+      "height": "${height}",
+      "batch_size": 1
+    }
+  },
+  "30": {
+    "class_type": "ModelSamplingFlux",
+    "inputs": {
+      "model": ["12", 0],
+      "max_shift": 1.15,
+      "base_shift": 0.5,
+      "width": "${width}",
+      "height": "${height}"
+    }
+  },
+  "31": {
+    "class_type": "KSampler",
+    "inputs": {
+      "model": ["30", 0],
+      "seed": "${seed}",
+      "steps": "${steps}",
+      "cfg": "${cfg}",
+      "sampler_name": "${sampler}",
+      "scheduler": "${scheduler}",
+      "denoise": 1.0,
+      "positive": ["6", 0],
+      "negative": ["13", 0],
+      "latent_image": ["27", 0]
+    }
+  }
+}
diff --git a/internal/backend/workflows/flux2-klein.json b/internal/backend/workflows/flux2-klein.json
new file mode 100644
index 0000000..1c97701
--- /dev/null
+++ b/internal/backend/workflows/flux2-klein.json
@@ -0,0 +1,79 @@
+{
+  "6": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "text": "${prompt}",
+      "clip": ["11", 0]
+    }
+  },
+  "8": {
+    "class_type": "VAEDecode",
+    "inputs": {
+      "samples": ["31", 0],
+      "vae": ["10", 0]
+    }
+  },
+  "9": {
+    "class_type": "SaveImage",
+    "inputs": {
+      "filename_prefix": "imagen",
+      "images": ["8", 0]
+    }
+  },
+  "10": {
+    "class_type": "VAELoader",
+    "inputs": {
+      "vae_name": "${vae}"
+    }
+  },
+  "11": {
+    "class_type": "CLIPLoader",
+    "inputs": {
+      "clip_name": "${clip}",
+      "type": "flux2"
+    }
+  },
+  "12": {
+    "class_type": "UNETLoader",
+    "inputs": {
+      "unet_name": "${model}",
+      "weight_dtype": "${dtype}"
+    }
+  },
+  "14": {
+    "class_type": "FluxGuidance",
+    "inputs": {
+      "conditioning": ["6", 0],
+      "guidance": "${guidance}"
+    }
+  },
+  "15": {
+    "class_type": "ConditioningZeroOut",
+    "inputs": {
+      "conditioning": ["6", 0]
+    }
+  },
+  "27": {
+    "class_type": "EmptyFlux2LatentImage",
+    "inputs": {
+      "width": "${width}",
+      "height": "${height}",
+      "batch_size": 1
+    }
+  },
+  "31": {
+    "class_type": "KSampler",
+    "inputs": {
+      "model": ["12", 0],
+      "seed": "${seed}",
+      "steps": "${steps}",
+      "cfg": "${cfg}",
+      "sampler_name": "${sampler}",
+      "scheduler": "${scheduler}",
+      "denoise": 1.0,
+      "positive": ["14", 0],
+      "negative": ["15", 0],
+      "latent_image": ["27", 0]
+    }
+  }
+}
diff --git a/internal/backend/workflows/sd35-medium.json b/internal/backend/workflows/sd35-medium.json
new file mode 100644
index 0000000..feb78a5
--- /dev/null
+++ b/internal/backend/workflows/sd35-medium.json
@@ -0,0 +1,66 @@
+{
+  "4": {
+    "class_type": "CheckpointLoaderSimple",
+    "inputs": {
+      "ckpt_name": "${model}"
+    }
+  },
+  "6": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "text": "${prompt}",
+      "clip": ["4", 1]
+    }
+  },
+  "7": {
+    "class_type": "CLIPTextEncode",
+    "inputs": {
+      "text": "${negative}",
+      "clip": ["4", 1]
+    }
+  },
+  "8": {
+    "class_type": "VAEDecode",
+    "inputs": {
+      "samples": ["31", 0],
+      "vae": ["4", 2]
+    }
+  },
+  "9": {
+    "class_type": "SaveImage",
+    "inputs": {
+      "filename_prefix": "imagen",
+      "images": ["8", 0]
+    }
+  },
+  "13": {
+    "class_type": "ModelSamplingSD3",
+    "inputs": {
+      "model": ["4", 0],
+      "shift": "${shift}"
+    }
+  },
+  "27": {
+    "class_type": "EmptySD3LatentImage",
+    "inputs": {
+      "width": "${width}",
+      "height": "${height}",
+      "batch_size": 1
+    }
+  },
+  "31": {
+    "class_type": "KSampler",
+    "inputs": {
+      "model": ["13", 0],
+      "seed": "${seed}",
+      "steps": "${steps}",
+      "cfg": "${cfg}",
+      "sampler_name": "${sampler}",
+      "scheduler": "${scheduler}",
+      "denoise": 1.0,
+      "positive": ["6", 0],
+      "negative": ["7", 0],
+      "latent_image": ["27", 0]
+    }
+  }
+}
diff --git a/internal/config/config.go b/internal/config/config.go
index a8a809e..4d4f236 100644
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -144,15 +144,54 @@ output:
   cloud_sync: auto
 
 backends:
+  # FLUX.1-schnell on the local ComfyUI server. The "workflow" key picks the
+  # bundled template under internal/backend/workflows/; omit it for back-compat
+  # (defaults to flux1-schnell). See docs/backends.md for the per-model setup.
   flux-schnell-local:
     type: comfyui
     base_url: http://mrock:8188
+    workflow: flux1-schnell
     # Filename of the unet checkpoint inside the ComfyUI server's
-    # models/unet/ directory. See docs/setup-comfyui-mrock.md.
+    # models/unet/ directory.
     model: flux1-schnell.safetensors
+    vae: ae.safetensors
+    clip_l: clip_l.safetensors
+    clip_t5: t5xxl_fp8_e4m3fn.safetensors
+    dtype: fp8_e4m3fn
     default_steps: 4
     default_sampler: euler
     default_scheduler: simple
+    default_cfg: 1.0
+
+  # FLUX.2 [klein] 4B distilled — sub-second on RTX 4070 Ti SUPER.
+  # Weights: BFL non-commercial; flux-2-klein-base-4b-fp8 in models/unet/,
+  # qwen_3_4b in models/text_encoders/, flux2-vae in models/vae/.
+  flux2-klein-local:
+    type: comfyui
+    base_url: http://mrock:8188
+    workflow: flux2-klein
+    model: flux-2-klein-base-4b-fp8.safetensors
+    vae: flux2-vae.safetensors
+    clip: qwen_3_4b.safetensors
+    dtype: fp8_e4m3fn
+    default_steps: 4
+    default_sampler: euler
+    default_scheduler: simple
+    default_cfg: 1.0
+    guidance: 4.0
+
+  # SD3.5 medium — single-checkpoint variant that bundles the three text
+  # encoders inside the .safetensors. Drop into models/checkpoints/.
+  sd35-medium-local:
+    type: comfyui
+    base_url: http://mrock:8188
+    workflow: sd35-medium
+    model: sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors
+    default_steps: 28
+    default_sampler: dpmpp_2m
+    default_scheduler: sgm_uniform
+    default_cfg: 4.5
+    shift: 3.0
 
   mock:
     type: mock