comfy-model-onboarding

comfy-model-onboarding

Use this skill to help configure model profiles for comfy-imagegen, comfy-videogen, and comfy-musicgen. Profiles are selected by capability, not by an entire tool. A profile is not an architecture: for example ltx23 is the architecture/adapter, while ltx23-10eros is one profile of that architecture.

For loose LoRAs, naming, moving, or selecting ad hoc LoRAs, use comfy-lora-onboarding. Do not add one-off style LoRAs to model profiles unless the user wants that LoRA to become a persistent default.

If comfy-models is not available, use comfy-tools-setup first. In this repository, prefer uv run comfy-models; outside the repo, let comfy-tools-setup install the CLIs with uv tool.

Setup

If .comfy-agent-tools.json is missing, ask the user where they want to store or use ComfyUI model files, then initialize and validate the local config:

uv run comfy-models init
uv run comfy-models set-models-dir <models_dir>
uv run comfy-models validate
uv run comfy-models show

If config already exists, respect its models_dir unless the user asks to change it.

Supported Architectures

Only configure these architectures in v1:

• qwen-image-edit: base profile qwen-edit2511, capabilities imagegen.generate, imagegen.edit.
• anima: base profile anima-base, capability imagegen.generate.
• flux-klein: base profile flux-klein-9b-snofs, capabilities imagegen.generate, imagegen.edit.
• ideogram4: base profile ideogram4-fp8, capability imagegen.ideogram4-generate.
• krea2: base profile krea2-turbo, optional local krea2-turbo-int4-fast profile, capability imagegen.krea2-generate.
• upscale-model: base profile clear-reality, capability imagegen.upscale.
• ltx23: base profile ltx23-10eros, optional Dasiwa profile ltx23-dasiwa-golden-lace-v3, capabilities videogen.t2v, videogen.i2v, videogen.flf2v, videogen.ia2av.
• wan22: base profiles wan22-t2v, wan22-i2v, wan22-s2v, and wan22-bernini, optional Dasiwa profiles wan22-dasiwa-tastysin-t2v, wan22-dasiwa-boundbite-t2v, wan22-dasiwa-tastysin-i2v, wan22-dasiwa-boundbite-i2v, wan22-dasiwa-littledemon-v2-s2v, and wan22-dasiwa-littledemon-v2-video-audio, capabilities videogen.wan22-t2v, videogen.wan22-i2v, videogen.wan22-flf2v, videogen.wan22-s2v, videogen.wan22-video-audio, videogen.wan22-bernini.
• seedvr2: base profile seedvr2, capability videogen.seedvr2-upscale.
• ace-step-1.5: base profile ace15-base, capability musicgen.generate.
• seedance2-api: remote profile seedance2-api, capabilities videogen.seedance2-t2v, videogen.seedance2-r2v, videogen.seedance2-flf2v.
• grok-imagine-api: remote profile grok-imagine-api, capabilities imagegen.grok-generate, imagegen.grok-edit.
• muapi-image-api: remote profile muapi-image-api, capability imagegen.muapi-generate.

If a model is SDXL, a non-WAN-2.2 Wan variant, a new audio architecture, or any Flux variant other than the built-in flux-klein adapter, do not pretend it is supported. Explain that a new architecture adapter is required.

Seedance 2.0 is not a local checkpoint or fine-tune architecture. Do not create local profiles for it, do not route it through models_dir, and do not use comfy-model-downloader. It requires COMFY_ORG_API_KEY and a comfy-diffusion version that vendors the ByteDance Seedance 2.0 API nodes.

Grok Imagine is also remote-only. Do not create local checkpoint profiles for grok-imagine-api, do not route it through models_dir, and do not use comfy-model-downloader. It requires COMFY_ORG_API_KEY and a comfy-diffusion version that vendors the Grok API nodes.

Ideogram 4 is local, not a remote API profile. It uses models_dir, supports download through comfy-model-downloader, and does not use COMFY_ORG_API_KEY. Use --unet, --uncond-unet, --clip, and --vae when adding compatible custom Ideogram 4 profiles.

SeedVR2 is local but its model files are managed by the pinned SeedVR2 upstream CLI, not by comfy-model-downloader or general models_dir onboarding. Use --models-dir/--model-dir on comfy-videogen seedvr2-upscale only when the user wants a specific SeedVR2 download/cache location.

Onboarding Flow

1. Inspect the user-provided path and current models dir with non-destructive

shell commands such as ls, find, or safetensors metadata reads.

1. Identify architecture first, then choose a base profile.
2. For fine-tunes, prefer extends and override only changed model paths.
3. Validate the profile before making it a default.
4. Set defaults per capability only where the profile supports that capability.
5. Recommend a small smoke test.

Fine-Tune Examples

LTX 2.3 checkpoint/fine-tune:

uv run comfy-models add-profile my-ltx23-finetune \
  --extends ltx23-10eros \
  --checkpoint checkpoints/my_ltx23_finetune.safetensors

uv run comfy-models validate-profile my-ltx23-finetune
uv run comfy-models set-default videogen.t2v my-ltx23-finetune

Dasiwa LTX 2.3 Golden Lace v3:

uv run comfy-models download-profile ltx23-dasiwa-golden-lace-v3 --dry-run
uv run comfy-models validate-profile ltx23-dasiwa-golden-lace-v3
uv run comfy-models set-default videogen.i2v ltx23-dasiwa-golden-lace-v3

WAN 2.2 Dasiwa T2V profile:

uv run comfy-models validate-profile wan22-dasiwa-tastysin-t2v
uv run comfy-models set-default videogen.wan22-t2v wan22-dasiwa-tastysin-t2v

Qwen Image Edit fine-tune:

uv run comfy-models add-profile my-qwen-edit \
  --extends qwen-edit2511 \
  --unet diffusion_models/my_qwen_edit.safetensors

uv run comfy-models validate-profile my-qwen-edit
uv run comfy-models set-default imagegen.generate my-qwen-edit
uv run comfy-models set-default imagegen.edit my-qwen-edit

Anima fine-tune or compatible LoRA:

uv run comfy-models add-profile my-anima \
  --extends anima-base \
  --unet diffusion_models/my_anima.safetensors \
  --lora loras/my_anima_turbo_lora.safetensors

uv run comfy-models validate-profile my-anima
uv run comfy-models set-default imagegen.generate my-anima

Anima profiles generate images only. They do not support imagegen.edit. Keep turbo Anima profiles at steps=8 and cfg=1.0 unless the user explicitly wants to test a slower non-turbo profile.

FLUX.2 Klein 9B SNOFS:

uv run comfy-models download-profile flux-klein-9b-snofs --dry-run
uv run comfy-models download-profile flux-klein-9b-snofs --yes
uv run comfy-models validate-profile flux-klein-9b-snofs
uv run comfy-models set-default imagegen.generate flux-klein-9b-snofs
uv run comfy-models set-default imagegen.edit flux-klein-9b-snofs

This profile uses gated Black Forest Labs FLUX.2 Klein 9B FP8 weights plus the SNOFS LoRA. Use HF_TOKEN after the user accepts the FLUX license terms. Do not redistribute the weights, and do not use this profile for a public/commercial generation service without separate licensing. Flux Klein edits follow the official distilled image-edit workflow from comfy-diffusion.

Ideogram 4 FP8:

uv run comfy-models download imagegen.ideogram4-generate --dry-run
uv run comfy-models download imagegen.ideogram4-generate --yes
uv run comfy-models validate-profile ideogram4-fp8

Ideogram prompts can be plain text, but structured JSON with bbox layout is preferred for typography and graphic design. Bboxes use y_min,x_min,y_max,x_max normalized 0..1000 coordinates.

ACE-Step 1.5 variant:

uv run comfy-models add-profile my-ace15 \
  --extends ace15-base \
  --unet diffusion_models/my_ace15.safetensors

uv run comfy-models validate-profile my-ace15
uv run comfy-models set-default musicgen.generate my-ace15

Smoke Tests

• Image generate: uv run comfy-imagegen generate --prompt "simple cinematic portrait" --width 512 --height 512
• Ideogram 4: uv run comfy-imagegen ideogram4-generate --prompt "minimal poster reading HELLO" --width 512 --height 512
• Image edit: use a small existing PNG and a simple prompt.
• Upscale: use a small existing PNG.
• Video: use --length 49 --fps 24.
• Music: use --duration 30 --steps 8 --cfg 1.

Do not install custom nodes or add unsupported adapters in this skill. For missing files in supported built-in profiles, use comfy-model-downloader. Keep changes scoped to .comfy-agent-tools.json and validated profile defaults.