Inpaint / Outpaint Workflow: style transfer from reference image with MLSD straight lines

You are designing an inpainting / outpainting workflow. This is surgical image editing: change a region while leaving the rest untouched, OR extend an image beyond its borders.

## Brief
- **Goal:** style transfer from reference image
- **Guidance:** MLSD straight lines
- **Base model:** AlbedoBase XL
- **Final upscaler:** APISR

## Inpaint vs Outpaint — Pick One Primary
- **Inpaint** = edit a region INSIDE the existing canvas. Use when: removing an object, replacing a face, changing wardrobe, cleaning up artifacts.
- **Outpaint** = extend the canvas OUTSIDE the original frame. Use when: converting vertical to horizontal, adding environment, zooming out.

## Workflow A — Inpainting

### Step 1. Mask creation
Two ways:
1. **Manual mask** — paint in the ComfyUI Load Image (as Mask) node or A1111 inpaint tab.
2. **Auto-mask** — use SAM (Segment Anything) with a text prompt: "face", "shirt", "sky".

Mask feathering: 5-10 px blur for smooth blending. Avoid hard edges — they create visible seams.

### Step 2. Model choice
- Use **inpaint-specific checkpoints** when possible: SDXL Inpainting, FluxInpaint. These are trained to respect mask boundaries.
- If using a non-inpaint AlbedoBase XL, apply the `DifferentialDiffusion` node to improve boundary quality.

### Step 3. Prompt
Describe ONLY what should be IN the masked region, in context.

Good: "a blue silk scarf wrapped around her neck, matching the lighting of the scene"
Bad: "a blue silk scarf" (the model doesn't know to match lighting)

### Step 4. Sampling
- **Denoise:** 0.7-0.95 for major changes; 0.3-0.5 for subtle edits
- **Steps:** 30
- **CFG:** 5-7
- **Conditioning mask strength:** 1.0 (full mask respect)
- **Mask blend with original:** via SetLatentNoiseMask or VAEEncodeForInpaint

### Step 5. ControlNet for inpaint
Add MLSD straight lines to guide the inpaint — for example, Depth ControlNet ensures new content respects the original scene depth.

### Step 6. Post-blend
Use `ImageCompositeMasked` to merge the inpainted region back into the original canvas with feathered mask.

## Workflow B — Outpainting

### Step 1. Canvas extension
Pad the image on the sides you want to extend. Typical padding: 256-512 px each side.

Use `ImagePadForOutpaint` node with:
- left, top, right, bottom: [pixels]
- feathering: 40-80 px

### Step 2. Prompt for the extended area
Describe the full extended scene. Include:
- Subject from original (anchor)
- Environment on the new edges
- Lighting continuity from original

### Step 3. Mask the new padded areas
The padded area becomes the inpaint mask automatically via the padding node.

### Step 4. Sampling
- Denoise: 1.0 on new area, 0.0 on original (full repaint of new area, zero on kept area)
- CFG: 5-7
- Steps: 30-40
- MLSD straight lines: use Tile or Depth extracted from the padded image to guide continuity

### Step 5. Second pass for seams
Sometimes the seam between original and outpainted area is visible. Run a second inpaint pass over a thin strip (30-50 px) along the seam with denoise 0.3-0.4 and Tile ControlNet to harmonize.

### Step 6. Final upscale
Use APISR to upscale the completed outpaint.

## Checkpoint-Specific Notes
- **SDXL Inpainting model:** Purpose-built. Use when available.
- **Flux.1 [dev] + FluxFill:** Flux's native inpainting; excellent quality.
- **Juggernaut XL v10 / RealVisXL:** Works with VAEEncodeForInpaint; use DifferentialDiffusion node.
- **SD 3.5 Large:** Emerging inpainting support.
- **Pony / Anime checkpoints:** Inpainting faces often drifts to default Pony face. Use FaceDetailer after.

## Pitfalls
- **Hard mask edges cause seams.** Always feather.
- **Low denoise + big change** = ghost of the original bleeds through. Raise denoise.
- **High denoise + small change** = over-shoot; use lower denoise + ControlNet.
- **Outpaint without continuity prompt** = discontinuous extension. Always describe the full scene, not just the new parts.
- **Upscaling a seamy outpaint** = upscales the seam. Fix seam BEFORE upscaling.

## Output
Return:

### 1. WORKFLOW TYPE
State: inpainting OR outpainting, based on style transfer from reference image.

### 2. NODE GRAPH
Complete ComfyUI graph for the chosen workflow, with values filled in.

### 3. PROMPTS
Positive + negative, tuned for AlbedoBase XL.

### 4. SETTINGS
All sampler, ControlNet, mask, and upscale values.

### 5. QA CHECKLIST
5 specific things the user should verify in the output.

### 6. ONE-LINE SUMMARY
"This workflow accomplishes style transfer from reference image by [approach]."

Generate the full deliverable.