ComfyUI Workflow Spec: generate architectural elevations from floor plan using Pony Diffusion XL

You are a ComfyUI workflow architect. Produce a complete, implementable node graph specification.

## Workflow Goal
- **Target outcome:** generate architectural elevations from floor plan
- **Base checkpoint:** Pony Diffusion XL
- **Primary control signal:** HED soft edge
- **Secondary technique:** IP-Adapter FaceID Portrait
- **Upscaling:** Topaz Gigapixel AI

## Output Structure
Deliver the workflow as four blocks:

### BLOCK 1 — Node Graph (textual representation)
List every node in execution order. For each node provide:
- Node type
- Inputs (upstream nodes)
- Key parameters
- Outputs (what it feeds)

Example formatting:
```
1. CheckpointLoaderSimple
   - ckpt_name: Pony Diffusion XL
   - outputs: MODEL, CLIP, VAE

2. CLIPTextEncode (positive)
   - text: [see positive prompt below]
   - clip: CLIP from node 1

3. CLIPTextEncode (negative)
   - text: [see negative prompt below]
   - clip: CLIP from node 1

[... continue for all nodes ...]
```

### BLOCK 2 — Required Custom Nodes
List every custom node pack the user must install:
- Name (as it appears in ComfyUI Manager)
- GitHub URL
- Purpose in this workflow
- Any model files to download separately (put in which folder)

Typical packs for this workflow:
- **ComfyUI-Impact-Pack** — FaceDetailer, SEGSDetailer
- **ComfyUI_IPAdapter_plus** — IP-Adapter integration
- **ComfyUI-Advanced-ControlNet** — ControlNet chaining
- **ComfyUI-AnimateDiff-Evolved** — if motion is needed
- **ComfyUI_UltimateSDUpscale** — high-res upscaling
- **ComfyUI-InstantID** — identity preservation

### BLOCK 3 — Prompts
- **Positive prompt:** [tuned for Pony Diffusion XL, 30-60 words]
- **Negative prompt:** [standard negatives + checkpoint-specific]
- **Prompt weighting strategy:** e.g., "(keyword:1.3)" for emphasis

### BLOCK 4 — Sampling Parameters
- KSampler settings:
  - Sampler: [e.g., DPM++ 2M Karras, Euler a, DDIM]
  - Scheduler: [karras, normal, exponential]
  - Steps: [20-40 typical]
  - CFG: [4-8 for SDXL; 3.5 for Flux; 5-7 for SD1.5]
  - Denoise: [1.0 for txt2img; 0.3-0.6 for img2img; 0.4-0.7 for ControlNet-guided]
- Second-pass settings (if HiRes fix or latent upscale used)
- ControlNet strength: [0.6-0.9 typical for HED soft edge]
- IP-Adapter weight: [0.5-0.8 for IP-Adapter FaceID Portrait]
- Upscale factor: [2x-4x with Topaz Gigapixel AI]

## Common ComfyUI Pitfalls
- **VRAM OOM:** Enable `--lowvram` or use tiled VAE decode. Pony Diffusion XL at 1024x1024 needs ~10 GB for SDXL.
- **Model mismatch:** HED soft edge models must match the base architecture (SD1.5 vs SDXL vs Flux). Download the matching ControlNet.
- **Noisy upscales:** Use Ultimate SD Upscale with 512px tiles and 32px padding; sets denoise to 0.15-0.3.
- **Identity drift:** Always use InstantID or IP-Adapter FaceID for character consistency.
- **Text in image:** ComfyUI's base Stable Diffusion cannot render text reliably — route that step through a Flux-in-ComfyUI node or composite externally.

## HED soft edge-Specific Notes
- **Canny:** Best for line-faithful stylization. Set low=100, high=200 thresholds.
- **Depth (MiDaS):** Best for composition transfer. Use "depth_zoe" for higher fidelity than MiDaS.
- **OpenPose:** Must use full body+face+hands model for consistent character poses.
- **Tile:** Essential for upscaling; chain with Topaz Gigapixel AI.
- **Normal:** Used with Depth for full 3D-awareness.
- **Scribble:** Low-fidelity rough sketches; 0.5 strength.
- **Segmentation:** Layout-level control; use ADE20K palette.

## IP-Adapter FaceID Portrait-Specific Notes
- **IP-Adapter Plus (style only):** Weight 0.5-0.8, use `weight_type: style transfer`.
- **IP-Adapter FaceID Plus v2:** Weight 0.7-0.9 for strong likeness; 0.4-0.5 for looser interpretation. Pair with InstantID for maximum identity lock.
- **IP-Adapter Composition:** Transfers composition only, not style or identity.

## Topaz Gigapixel AI-Specific Notes
- **4x-UltraSharp:** Fastest, slight smoothing. Good default.
- **SUPIR:** Highest quality, slowest. Use for final hero shots. Run at 2x, not 4x, to avoid OOM.
- **Topaz Gigapixel:** External tool; use after ComfyUI export.
- **Real-ESRGAN x4plus:** Cartoon/anime friendly.

## Deliver
Return all four blocks, filled with specific values for this workflow. Keep the spec implementation-ready — no hand-waving. The user should be able to build this in ComfyUI tonight.