Extensions/OnlyGemini
ComfyUI Extension

OnlyGemini

ComfyUI custom nodes wrapping Google Gemini APIs for image generation, multi-image composition, video generation (Veo and Gemini Omni Flash), and local LaMa inpainting.

By choquevolqueta·Created 6 days ago·Updated a day ago· 0
choquevolqueta/OnlyGemini
Nodes—
On cloudLocal install
Stars0
Updateda day ago
Readme

OnlyGemini — ComfyUI Custom Nodes

License Python ComfyUI

A small, focused pack of ComfyUI custom nodes that wrap Google Gemini APIs for image generation, multi-image composition, and video generation — plus one fully local (offline) inpainting node.

5 nodes: 4 use the Gemini API (google-genai), 1 is 100% local (LaMa, no internet).

Features

| Node | Display name | Category | API? | What it does | |------|--------------|----------|------|--------------| | NanoBananaGen | OG NanoBanana Generate | image/generation | ✅ Gemini | Text-to-image and image editing (img2img), up to 5 reference images | | LaMaRefill | OG LaMa Refill | image/editing | ❌ Local | Inpainting: remove objects with a mask (LaMa, no internet) | | GeminiOrchestrate | OG Gemini Orchestrate | image/generation | ✅ Gemini | Compose one scene from a background + up to 5 references | | VeoGenerate | OG Veo Generate | video/generation | ✅ Gemini | Veo video generation (text2video, first+last, ingredients, extension) | | OmniFlashGen | OG Omni Flash | video/generation | ✅ Gemini | Edit/restyle an existing video with Gemini Omni Flash + up to 6 reference images |

Requirements

  • ComfyUI (recent version with the native VIDEO type for OG Omni Flash).
  • Python 3.10+.
  • A Gemini API key from Google AI Studio.
  • OG LaMa Refill downloads the LaMa model (~200 MB) to disk on first use; afterwards it runs entirely offline.

Installation

Via ComfyUI Manager (recommended)

  1. Open ComfyUI Manager → Install Custom Nodes.
  2. Search for "OnlyGemini" and install.
  3. Restart ComfyUI.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/Choquevolqueta/OnlyGemini.git
cd OnlyGemini
pip install -r requirements.txt

API key setup

Three options, in order of preference:

  1. Environment variable (recommended): 
    export GEMINI_API_KEY="your-api-key-here"
  2. .env file in the pack folder (git-ignored): 
    GEMINI_API_KEY=your-api-key-here
  3. Per-node: type the key directly into each node's api_key field.

api_key also accepts the name of an environment variable. Connect a text widget with the value GEMINI_API_KEY and the pack resolves it automatically.

Node reference

All display names are prefixed with OG so they group together in the node search.

1. OG NanoBanana Generate (NanoBananaGen)

Text-to-image and image editing via generate_content.

  • Text-to-image: connect only prompt.
  • Image editing: connect image + a prompt describing the change.
  • Reference guidance: connect ref_image_1..5.

Inputs: prompt, api_key, model, aspect_ratio, resolution (only gemini-3-pro models), temperature, top_p, seed (-1 = random), optional system_instruction, google_search, image, ref_image_1..5.

Output: image — tensor [B, H, W, C] float32 0–1.

2. OG LaMa Refill (LaMaRefill)

Local, offline inpainting. Paint the object to remove in the MaskEditor, connect image + mask.

Inputs: image, mask, dilate_px, feather_px, passes, context_padding.

Output: image with the masked object removed.

3. OG Gemini Orchestrate (GeminiOrchestrate)

Composes a final scene from a required background and up to 5 reference images (subjects/objects). References are resized to the background automatically.

Inputs: prompt, api_key, model, aspect_ratio, resolution, temperature, top_p, seed, system_instruction, background (required), ref_image_1..5.

Output: image.

4. OG Veo Generate (VeoGenerate)

Veo video generation. The mode is detected automatically from what you connect:

| Mode | Connect | |------|---------| | Text-to-video | only prompt | | First + last frame | image + last_image | | Ingredients | reference_images (up to 3) | | Extension | video_uri from a previous Veo run |

Inputs: prompt, model (veo-3.1-generate-preview / -fast / -lite), aspect_ratio, duration_seconds (4–8), resolution, api_key, seed, optional negative_prompt, image, last_image, reference_images, video_uri.

Outputs: frames (list), audio, video_uri.

Notes

  • seed is exposed for workflow compatibility but is not sent to the API — the Gemini Developer API does not support a seed for Veo video.
  • video_uri only accepts a Veo-generated video (720p, referenced within 2 days); it cannot take an arbitrary file.

5. OG Omni Flash (OmniFlashGen)

The pack's video-editing node, built on Gemini Omni Flash (gemini-omni-1.1-flash) via the Interactions API. Unlike Veo, it can take your own video and restyle it while following its motion.

Typical use: render a previsualization in Blender (camera movement + character motion), then connect it together with 4–6 aesthetic reference images and a prompt. The node uploads the video, edits it, and returns the restyled frames.

| task | What it does | |--------|--------------| | edit | Edit/restyle an input video (requires video) | | text_to_video | Generate from text | | image_to_video | Animate an image | | reference_to_video | Generate from reference images | | auto | Let the model infer |

Inputs: prompt, model, task, aspect_ratio, resolution, seed (Omni does support a seed), api_key, optional video (VIDEO), ref_image_1..6 (IMAGE), first_frame, last_frame, previous_interaction_id.

Outputs: frames (list), audio, interaction_id (to chain edits).

Limits / caveats:

  • Input video for editing is 10 s max; longer clips are trimmed to the first 10 s.
  • Output is 3–10 s, 720p native (24 fps); 1080p/4k are upscaled.
  • Uploaded-video editing is not available in the EEA / Switzerland / UK.
  • Media containing recognizable people or minors is blocked by safety filters.
  • The result is not frame-perfect: the model regenerates guided by the input motion, so camera/action are preserved approximately, not pixel-exactly.

Example workflow: Blender previsualization → styled video

  1. Blender: render a short (≤10 s) previsualization — the layout, camera path, and character motion you want to keep.
  2. ComfyUI: Load Video → the video input of OG Omni Flash.
  3. Generate or load 4–6 style references (e.g. OG NanoBanana Generate outputs) → ref_image_1..6.
  4. Set task = edit and a prompt like "Restyle this video to match the reference look; keep the camera movement and character motion."
  5. Run. frames/audio carry the result; interaction_id lets you iterate with previous_interaction_id.

Troubleshooting

  • API key required — no key found. Set GEMINI_API_KEY or create .env.
  • 400 / invalid argument on video — Veo extension only accepts Veo-generated videos (720p). For arbitrary video, use OG Omni Flash.
  • Omni edit rejected — your video may be >10 s, in an unsupported region, or contain people/minors.
  • [Errno 11001] getaddrinfo failed / timeouts — transient network/DNS; retry.
  • LaMa slow on first run — it is downloading the model; later runs use the cache.

Limitations

  • English prompts are fully supported; other languages may vary in quality.
  • Video generation is asynchronous and can take minutes under load.
  • Generated videos are SynthID-watermarked.

Development

# Install dev tools
pip install -r requirements-dev.txt

# Lint / format
ruff check .
ruff format .

# Tests (standalone scripts; also runnable under pytest)
python tests/test_utils.py
python tests/test_omni_gen.py

The node modules live in nodes/ and share helpers in utils.py. To add a node: create nodes/<name>.py, export NODE_CLASS_MAPPINGS + NODE_DISPLAY_NAME_MAPPINGS, and add the module name to _node_modules in __init__.py.

License

MIT © 2026 Alexander Cortez.