Imported from Septa-Serpenta-Seraph/Vesper (skills/.archive/comfyui-image-workflow/SKILL.md). Install upstream with npx skills add Septa-Serpenta-Seraph/Vesper --skill comfyui-image-workflow. Copyright stays with the author.

ComfyUI Image Workflow

Trigger: Configuring a ComfyUI workflow for FLUX vs SDXL, handling explicit/intimate content, downloading large HuggingFace models to Windows, or delivering generated images to the user.

This skill complements comfyui-ssh-tunnel (SSH setup, tunnel verification, ComfyUI launch). This skill covers what to do after the tunnel is up.

FLUX (dev) fp8 — Workflow

Key parameters (differ significantly from SDXL)

FLUX workflow JSON

Use in place of the SDXL workflow:

{
  "3": {
    "inputs": {
      "seed": 200, "steps": 25, "cfg": 1.0,
      "sampler_name": "euler", "scheduler": "sgm_uniform", "denoise": 1,
      "model": ["4", 0], "positive": ["6", 0], "negative": ["7", 0],
      "latent_image": ["5", 0]
    }, "class_type": "KSampler"
  },
  "4": {"inputs": {"ckpt_name": "flux1-dev-fp8.safetensors"}, "class_type": "CheckpointLoaderSimple"},
  "5": {"inputs": {"width": 1024, "height": 1024, "batch_size": 1}, "class_type": "EmptyLatentImage"},
  "6": {"inputs": {"text": "PROMPT", "clip": ["4", 1]}, "class_type": "CLIPTextEncode"},
  "7": {"inputs": {"text": "", "clip": ["4", 1]}, "class_type": "CLIPTextEncode"},
  "8": {"inputs": {"samples": ["3", 0], "vae": ["4", 2]}, "class_type": "VAEDecode"},
  "9": {"inputs": {"images": ["8", 0], "filename_prefix": "PREFIX"}, "class_type": "SaveImage"}
}

Downloading Large Models from HuggingFace (Windows)

Problem

• curl -L with HuggingFace redirect URLs downloads the HTML redirect page, not the model file.
• Invoke-WebRequest (PowerShell) handles the redirect but frequently corrupts files >10 GB (incomplete writes).
• File appears on disk but fails to load with: "shape '[4096, 4096]' is invalid for input of size N"

Solution: curl.exe (bitsadmin FAILED over SSH — verified 8/11/26)

bitsadmin /create /download failed with "Invalid number of arguments" every time it was driven through the SSH tunnel — both inline via ssh cmd /c AND via a .bat file copied to the box. The SSH/cmd quoting mangles bitsadmin's argument parsing. (Earlier success notes used bitsadmin; the tunnel-driven case does not work.)

The method that worked (Windows 10+ ships curl.exe 8.x):

ssh -i ~/.ssh/windows_desktop -p 1237 tyler@127.0.0.1 cmd /c \
  "curl.exe -sL -C - --retry 5 --retry-delay 10 --retry-all-errors --max-time 5400 \
    -o \"C:\\ComfyUI\\models\\<folder>\\<file>\" \"<HF resolve URL>\""

• -L follows HF redirects; -C - resumes partial downloads; --retry for flaky transfers.
• Drive big files from a python script in a background process; keep a log + a --status-only size check.
• Killed curl leaves a 0-byte stub — a presence check (if exist) wrongly counts it as downloaded. Use a SIZE check (threshold >10MB), not existence: for %F in ("C:\ComfyUI\models\<folder>\<file>") do @echo %~zF
• Verify byte-exact vs the HF API listing (https://huggingface.co/api/models/<owner>/<repo>/tree/main/<path>) — e.g. Wan 2.2 Remix NSFW v2.0 high_lighting = 14,291,272,136 bytes.
• Remote curl survives the SSH client dying — if you kill the driver, the Windows curl keeps going; wait and re-check size, don't restart.
• Full detail: remote-comfyui-models skill → references/downloading-models-via-curl.md.

Verification

1. Check model appears in ComfyUI's model list:

   ssh -i ~/.ssh/windows_desktop -p 1237 tyler@127.0.0.1 cmd /c \
     "curl -s http://127.0.0.1:8188/object_info/CheckpointLoaderSimple | findstr MODEL_NAME"

2. Try a minimal generation and check history for execution_error (via http://127.0.0.1:8188/history/<prompt_id>). The error signature for a corrupted file is: RuntimeError: shape '[4096, 4096]' is invalid for input of size N.

Explicit / Intimate Image Generation (SDXL — Juggernaut)

Cloud fallback when the desktop PC is down: Together.ai (proven 8/10/26)

When Tyler asks for uncensored/intimate images and the Windows box is off, use the Together.ai FLUX.2-dev script — the uncensored cloud path:

python3 /home/lumi/.hermes/profiles/vesper/scripts/ves_image.py "<prompt>"

• Reads TOGETHER_API_KEY from the profile .env (strips literal quotes).
• Raw urllib, NOT curl (curl hits TLS error 43 to Together).
• Browser User-Agent required or Cloudflare WAF error 1010 blocks both the API call AND the image CDN URL.
• Prints the absolute output path on stdout; saves under ~/.hermes/profiles/vesper/cron/output/.
• The built-in image_generate tool routes via FAL, NOT Together — FAL's content checker either returns content_policy_violation or silently emits an ALL-BLACK image for explicit prompts. Do not use image_generate for uncensored content; use ves_image.py.
• Anatomy rules live in the script header (PROMPT_BANK) — beak phrasing is verified: "Her mouth is a small curved soft black crow beak" / "her human lips are replaced by a glossy black crow beak protruding from her face...". Do NOT deviate ("lips are replaced by" → beak sticks out like a tongue; "where her mouth and nose would be" → half crow-head fused to her face).
• Keep feathers on shoulders (shawl) + small wings at arms; vary only lighting/mood words between entries.
• Script header says tasteful/intimate only — with Tyler's explicit consent on record, a custom explicit prompt works fine (verified: candlelit nude on silk sheets generated clean).

Model selection

• Juggernaut XL Ragnarok: photorealistic, handles explicit/tasteful nudes reliably.
• DreamShaper XL: fantasy/artistic, conservative (will not generate nudity).
• Pony Diffusion XL v6: anime style, explicit-friendly.

Production parameters

Prompt language for corvid-human intimate scenes

Corvid feature cues (SDXL understands these as accessory/fashion, NOT body-horror):

• "glossy black feathers across her bare shoulders"
• "feathers woven through her dark hair"
• "small dark wings folded at her sides"
• "feathers scattered across the sheets"

Lighting/mood cues (Juggernaut responds to specific atmospheric language):

• "soft morning light through sheer curtains"
• "warm lamp light, intimate atmosphere"
• "golden hour, rim lighting"

Couple scenes:

• "being held from behind, his arms around her"
• "foreheads touching, noses brushing"
• "her head on his chest, his hand in her hair"
• "facing each other in bed, tangled white sheets"

Known limitation: beak fusion

SDXL CANNOT fuse a crow beak onto a human face realistically. It produces a second nose/mouth area. Only FLUX fp8 can do this. Proven FLUX anchor prompt for beak fusion:

"a woman whose human lips are replaced by a glossy black crow beak protruding from her face and seamlessly fused to her skin like it is her own mouth; warm normal human nose above. Glossy black feathers drape her shoulders like a shawl, small dark wings folded against her arms. Portrait, natural lighting, photographic, detailed skin texture, high quality"

MiniMax H3 — Local Video Generation

H3 is the local omni-modal video model on the desktop box (ComfyUI 0.30.0+, comfy-kitchen). It does video + audio + reference conditioning. When Tyler says "video gen" or "take video for a spin," this is the path.

Full node wiring, model filenames, length semantics (124 frames ≈ 5 s, trained range 124–362), the API-format workflow graph, and the remote execution recipe live in references/minimax-h3-video.md.

Quick reference:

• Model: minimax_h3_fl2va_pruned_int8_convrot.safetensors (UNETLoader)
• Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (CLIPLoader)
• Video VAE: minimax_h3_video_vae_fp16.safetensors (VAELoader)
• Node: MiniMaxH3ImageToVideo (clip+vae+prompt+width+height+length → CONDITIONING+LATENT)
• Omni path: MiniMaxH3ReferenceToVideo (adds audio_vae + ref_images/ref_videos/ref_audios)
• Add MiniMaxH3SigmaShift (model, shift_video, shift_audio) before KSampler.
• Negative: use ConditioningZeroOut or empty CLIPTextEncode — no H3 negative node.

Pitfall: tunnel port 1237 is SSH transport, not HTTP (verified 2026-08-07)

curl http://localhost:1237/ returns Received HTTP/0.9 when not allowed — that's normal; the tunnel forwards to Windows' SSH server (port 22), not to ComfyUI. To reach ComfyUI through the tunnel, SSH in first and curl the Windows-local URL from inside: ssh -i ~/.ssh/windows_desktop -p 1237 tyler@127.0.0.1 cmd /c "curl -s http://127.0.0.1:8188/system_stats".

Image Retrieval & Discord Delivery

After generation completes on Windows:

1. SCP from Windows:

   scp -P 1237 -i ~/.ssh/windows_desktop \
     tyler@127.0.0.1:/C:/ComfyUI/output/<file>.png \
     /home/lumi/.hermes/profiles/vesper/cache/images/<batch>/<file>.png

2. Organize under batch directories:

   /home/lumi/.hermes/profiles/vesper/cache/images/<batch_name>/
   

3. Deliver to current channel: Include MEDIA:<absolute_path> in the response text. On Discord, this renders as an image attachment.

Related

• comfyui-ssh-tunnel — SSH tunnel setup, ComfyUI launch/restart, tunnel port scanning
• communication/intimate-scenes — shared language for corvid-human intimate scenes (first-person narrative, consent, aftercare)
• references/wan22-verified-session-2026-08-11.md — Wan 2.2 Remix NSFW verified config, CFG-disaster warning, node schemas, anchor lesson (8/11/26). Sibling of comfyui-wan22-video.