Comfy Imagegen

comfy-imagegen

Use this skill for image generation, image editing, and image upscaling through the comfy-imagegen CLI. Local modes use models under /mnt/models/comfyui. If a supported built-in model is missing, use comfy-model-downloader to fetch only the requested capability before running inference.

The CLI is quiet by default and prints only final JSON. Use --verbose only when debugging ComfyUI runtime output, warnings, or progress bars.

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

At the start of every image workflow, start or reuse the local Comfy Media gallery for the active output directory:

uv run comfy-media gallery --out outputs --host 127.0.0.1 --port 8765

Use comfy-media --help only if the CLI is missing or behaves unexpectedly; do not skip the gallery just because generation can run headless.

If .comfy-agent-tools.json is missing or the user wants to configure a new checkpoint/fine-tune/default, use comfy-model-onboarding first.

If model validation fails with missing_model_file, use comfy-model-downloader: imagegen.generate for the active generation profile, imagegen.edit for the active edit profile, imagegen.ideogram4-generate for Ideogram 4, imagegen.krea2-generate for Krea2 Turbo, or imagegen.upscale for ClearReality upscale.

If the user asks to use or organize a LoRA by name or purpose, use comfy-lora-onboarding to search loras// first and pass the chosen file with --extra-lora.

Modes

• generate: text prompt to image with Anima Base v1.0 + turbo LoRA. Use for

anime, illustration, and non-photorealistic art by default. When the active profile is flux-klein-9b-snofs, use FLUX.2 Klein 9B FP8 plus SNOFS.

• edit: input image plus prompt to edited image with Qwen Image Edit 2511.

When the active profile is flux-klein-9b-snofs, use FLUX.2 Klein 9B plus SNOFS for single-reference editing.

• upscale: input image to 4x upscale with ClearReality.
• rtx-upscale: input image to a target resolution (480p/720p/1080p/1440p/4k/8k),

custom width/height, or 1.0-4.0 scale with NVIDIA RTX Video Super Resolution. Requires an NVIDIA RTX GPU and the nvidia-vfx package.

• ideogram4-generate: local Ideogram 4 text-to-image. Use this for poster,

typography, graphic design, color palette, and bbox/layout-controlled images.

• krea2-generate: local Krea2 Turbo text-to-image. Use this for fast, high-

fidelity prompt-following generation with a Qwen3-VL text encoder and the Krea2 conditioning rebalance.

• grok-generate: remote Grok Imagine text prompt to PNG through Comfy API

nodes. Requires COMFY_ORG_API_KEY, not local model files.

• grok-edit: remote Grok Imagine input image plus prompt to PNG through Comfy

API nodes. Requires COMFY_ORG_API_KEY, not local model files.

Commands

Generate:

uv run comfy-imagegen generate \
  --prompt "masterpiece, best quality, score_7, safe, 1girl, anime style, cinematic lighting, detailed background" \
  --width 1024 \
  --height 1024 \
  --seed 42 \
  --extra-lora /mnt/models/comfyui/loras/anima/realism-portrait.safetensors:0.8:0.0 \
  --out outputs

Edit:

uv run comfy-imagegen edit \
  --input path/to/input.png \
  --prompt "Transform this photo into a polished animated film still style while preserving the main subject and composition" \
  --seed 43 \
  --out outputs

For FLUX.2 Klein edit profiles, pass --width and/or --height to choose the output canvas. Any omitted dimension defaults to the input image dimension.

Upscale:

uv run comfy-imagegen upscale \
  --input path/to/input.png \
  --out outputs

RTX upscale to a target resolution:

uv run comfy-imagegen rtx-upscale \
  --input path/to/input.png \
  --resolution 4k \
  --quality ULTRA \
  --out outputs

RTX upscale by scale factor or custom size (requires an NVIDIA RTX GPU and nvidia-vfx):

uv run comfy-imagegen rtx-upscale --input path/to/input.png --scale 2.0 --out outputs
uv run comfy-imagegen rtx-upscale --input path/to/input.png --width 2560 --height 1440 --out outputs

Use exactly one of --resolution, --scale, or --width/--height. The aspect ratio is preserved for --resolution and --scale; --width and --height must be provided together. Quality presets are LOW, MEDIUM, HIGH, and ULTRA (default ULTRA).

Ideogram 4:

uv run comfy-imagegen ideogram4-generate \
  --prompt "A poster for a jazz night" \
  --style-aesthetics "minimal, geometric, high contrast" \
  --style-lighting "flat graphic design lighting" \
  --style-medium "graphic_design" \
  --style-art-style "clean vector poster, sans-serif typography" \
  --background "Matte black paper background with subtle grain." \
  --text "90,120,260,880|JAZZ NIGHT|Large cream uppercase headline." \
  --output-json outputs/ideogram4-prompt.json \
  --width 1024 \
  --height 1024 \
  --out outputs

Ideogram 4 bbox builder:

uv run comfy-imagegen ideogram4-generate \
  --prompt "A modern concert poster for a jazz trio." \
  --style-aesthetics "minimal, elegant, high contrast" \
  --style-lighting "flat graphic design lighting" \
  --style-medium "graphic_design" \
  --style-art-style "clean vector poster, sans-serif typography" \
  --style-color "#101010" \
  --style-color "#F4D35E" \
  --background "Solid black background with subtle paper texture." \
  --object "420,120,900,880|A golden saxophone centered in the lower half." \
  --text "80,120,220,880|JAZZ NIGHT|Large condensed yellow headline." \
  --output-json outputs/ideogram4-poster-prompt.json \
  --out outputs

Krea2 Turbo:

uv run comfy-imagegen krea2-generate \
  --prompt "a cinematic portrait of an astronaut floating in a nebula, dramatic rim light" \
  --width 1024 \
  --height 1024 \
  --steps 8 \
  --cfg 1 \
  --rebalance-multiplier 4.0 \
  --out outputs

Grok Imagine:

COMFY_ORG_API_KEY=... uv run comfy-imagegen grok-generate \
  --prompt "A cinematic product photo of a translucent orange cassette player on wet asphalt" \
  --model grok-imagine-image \
  --resolution 1K \
  --aspect-ratio 1:1 \
  --out outputs

Grok Imagine edit:

COMFY_ORG_API_KEY=... uv run comfy-imagegen grok-edit \
  --input path/to/input.png \
  --prompt "Keep the subject and composition, change the background to a clean moonlit studio" \
  --resolution 1K \
  --out outputs

Prompt Guidance

For generation, Anima accepts Danbooru-style tags, natural language, or a mix. Use lowercase tags with spaces instead of underscores except score tags. Recommended positive prefix: masterpiece, best quality, score_7, safe, .... Recommended negative guidance when adapting prompts: worst quality, low quality, score_1, score_2, score_3, artist name.

Anima is not a realism model. It is intended for anime, illustration, and art. Keep generation around 1MP, such as 1024x1024, 896x1152, or 1152x896.

For editing, write an instruction that names what should change and what should stay fixed. Preserve identity, layout, pose, and important object details unless the user explicitly asks to change them.

Qwen Image Edit may rescale internally, so read outputs[].width and outputs[].height from the final JSON instead of assuming requested or input dimensions match the saved PNG.

For upscale, do not add prompt text; use the input image only.

Ideogram 4 is trained for structured JSON captions, and the CLI builds that JSON from flags. Do not pass raw JSON. Required fields are --prompt, --style-aesthetics, --style-lighting, --style-medium, --background, exactly one of --style-photo or --style-art-style, and at least one --object or --text. Bboxes are y_min,x_min,y_max,x_max in normalized 0..1000 coordinates, not pixels. Use --object "BBOX|DESCRIPTION" for visual subjects and --text "BBOX|TEXT|DESCRIPTION" for literal rendered text. Use uppercase #RRGGBB color values; the CLI normalizes --style-color values to uppercase. Use --output-json PATH when the generated Ideogram prompt JSON should be saved for inspection or reuse.

Grok Imagine uses remote Comfy API nodes and is separate from local model profiles. Do not route grok-imagine-api through models_dir, comfy-model-downloader, local LoRAs, or checkpoint onboarding. Supported image models are grok-imagine-image-pro, grok-imagine-image, and grok-imagine-image-beta; supported resolutions are 1K and 2K.

FLUX.2 Klein 9B SNOFS uses natural-language prompts, supports both generation and editing, and is step-distilled for steps=4, cfg=1.0. Keep dimensions divisible by 16. It requires gated Black Forest Labs weights and SNOFS has a separate personal-use license: local image generation is allowed, generated images may be sold, but public/commercial generation services, derivative model creation, and weight redistribution are not allowed without a separate license. For edits, Flux Klein follows the official distilled image-edit workflow: reference-image VAE encoding, reference latents on both positive and negative conditioning, Flux2Scheduler, CFGGuider, and SamplerCustomAdvanced.

Defaults

• Models directory: /mnt/models/comfyui
• Generate profile: anima-base
• Anima diffusion model: diffusion_models/anima-base-v1.0.safetensors
• Anima text encoder: text_encoders/qwen_3_06b_base.safetensors
• VAE: vae/qwen_image_vae.safetensors
• Anima turbo LoRA: loras/anima/anima-turbo-lora-v0.1.safetensors
• Anima params: steps=8, cfg=1.0, seed=0
• Edit profile: qwen-edit2511
• Qwen diffusion model: diffusion_models/qwen_image_edit_2511_fp8mixed.safetensors
• Qwen text encoder: text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors
• Qwen LoRA: loras/qwen-image-edit/Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors
• Qwen params: steps=4, cfg=3.0, seed=0
• FLUX profile: flux-klein-9b-snofs
• FLUX architecture: flux-klein
• FLUX diffusion model: diffusion_models/flux-2-klein-9b-fp8.safetensors
• FLUX text encoder: text_encoders/qwen_3_8b_fp8mixed.safetensors
• FLUX VAE: vae/flux2-vae.safetensors
• SNOFS LoRA: loras/flux-klein/klein_snofs_v1_1.safetensors
• FLUX params: steps=4, cfg=1.0, sampler=euler, seed=0
• Ideogram profile: ideogram4-fp8
• Ideogram diffusion model: diffusion_models/ideogram4_fp8_scaled.safetensors
• Ideogram unconditional model: diffusion_models/ideogram4_unconditional_fp8_scaled.safetensors
• Ideogram text encoder: text_encoders/qwen3vl_8b_fp8_scaled.safetensors
• Ideogram VAE: vae/flux2-vae.safetensors
• Ideogram params: steps=20, cfg=7.0, cfg_override_value=3.0, sampler=euler, seed=0
• Krea2 profile: krea2-turbo
• Krea2 diffusion model: diffusion_models/krea2_turbo_fp8_scaled.safetensors
• Krea2 text encoder: text_encoders/qwen3vl_4b_fp8_scaled.safetensors
• Krea2 VAE: vae/qwen_image_vae.safetensors
• Krea2 params: steps=8, cfg=1.0, sampler=euler, scheduler=simple, rebalance_multiplier=4.0, seed=0
• Upscaler: upscale_models/4x-ClearRealityV1.pth
• RTX upscale profile: rtx-vsr (shared with videogen.rtx-upscale)
• RTX upscale params: resolution=1080p, quality=ULTRA; presets 480p, 720p, 1080p, 1440p, 4k, 8k; --scale 1.0-4.0
• RTX upscale dependency: nvidia-vfx plus a CUDA-capable NVIDIA RTX GPU
• Grok profile: grok-imagine-api
• Grok provider: comfy-api
• Grok model: grok-imagine-image
• Grok params: resolution=1K, aspect_ratio=1:1, number_of_images=1, seed=0
• Dependency: comfy-diffusion[video,audio] v2.4.5 or newer plus the vendored

ComfyUI requirements

The Anima turbo LoRA expects cfg=1.0; increasing CFG can degrade or break the expected turbo behavior.

Extra LoRAs are optional and ad hoc. Use repeatable --extra-lora PATH[:MODEL_STRENGTH[:CLIP_STRENGTH]] after resolving the file through the LoRA folder convention:

• loras/anima/ for Anima generation LoRAs.
• loras/qwen-image-edit/ for Qwen edit LoRAs.
• loras/flux-klein/ for FLUX.2 Klein LoRAs such as SNOFS.
• loras/ideogram4/ for Ideogram 4 style or poster/detail LoRAs.
• fallback to loose files in loras/ only when no architecture folder match is clear.

For Ideogram 4, --extra-lora patches both the main and unconditional UNets. Nonzero CLIP_STRENGTH patches the shared Ideogram text encoder once.

Warnings about missing torchaudio can appear in verbose mode because ComfyUI imports audio-capable modules. They are expected while this skill is only using image workflows.

Output Handling

The CLI prints JSON to stdout. On success, read artifacts for the saved PNG paths and outputs for final image dimensions. On failure, read error and error_type; do not parse logs for control flow.

Project-bound images should be saved under the current workspace. Do not overwrite existing assets unless the user explicitly asks for replacement.

After every successful image generation, edit, or upscale command, immediately index the same output directory so the new artifact appears in Comfy Media:

uv run comfy-media index --out outputs

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: quinteroac
• Source: quinteroac/comfy-agent-tools
• License: MIT

Install and usage instructions live in the source repository linked above.