Grok Vision Analyze — send an image (tensor or URL) to xAI's Grok vision API and get a text response back. Useful for auto-captioning, prompt generation from reference images, or quick visual QA.
Load Image From URL — fetch a remote image directly into ComfyUI as IMAGE + MASK tensors. No download/upload step needed.
SCAIL-2 Infinity (auto window) — generate arbitrarily long SCAIL-2 character animation from one node, instead of hand-wiring chunk after chunk.

Install

Clone into your ComfyUI custom_nodes folder:

cd ComfyUI/custom_nodes
git clone https://github.com/DanielBartolic/ComfyUI-DioBrando-Nodes
cd ComfyUI-DioBrando-Nodes
pip install -r requirements.txt   # only Pillow + numpy + torch, all standard ComfyUI deps

Restart ComfyUI. The nodes show up under the DioBrando/ category.

Grok Vision setup

Set your xAI API key as an environment variable before launching ComfyUI:

export XAI_API="xai-..."

Or paste it into the optional api_key input on the node (overrides env var).

Models supported

grok-4.3-latest          ← default
grok-4.3
grok-4.20-multi-agent-0309
grok-4.20-0309-reasoning
grok-4.20-0309-non-reasoning
grok-4-1-fast-reasoning      ← fastest + cheapest
grok-4-1-fast-non-reasoning  ← fastest + cheapest
grok-4-vision
grok-3-vision-beta
grok-2-vision-1212
grok-vision-beta

SCAIL-2 Infinity

SCAIL-2 is trained on 81-frame chunks with a 5-frame overlap (76-frame step). Going past 81 frames normally means wiring multiple WanSCAILToVideo → KSampler → VAEDecode blocks by hand, feeding each chunk's decoded tail into the next via previous_frames / video_frame_offset. This node runs that loop internally — chunk → sample → decode → re-anchor on the last 5 frames → repeat until the driving pose video is exhausted — and stitches the result into one continuous video.

Nothing is reimplemented: it calls core WanSCAILToVideo.execute() for per-chunk conditioning, common_ksampler() to sample, and VAE.decode / decode_tiled to decode. All SCAIL-2 features (pose mask, reference mask, replacement mode, clip-vision, pose strength/start/end) pass straight through. Peak VRAM stays at one 81-frame window regardless of total length — each decoded chunk moves to CPU immediately and the cache is freed per window.

| Param | Default | Notes | |---|---|---| | window_length | 81 | Frames per chunk. SCAIL-2 trained at 81 — keep it. | | previous_frame_count | 5 | Overlap frames anchored from the previous chunk. Trained at 5. | | max_frames | 0 | Hard cap on total frames. 0 = run until the pose video ends. | | decode_tiled | off | Tiled VAE decode, to bound decode VRAM at high resolution. | | vary_seed_per_window | off | Off = same seed each chunk (best continuity). |

With a pose_video it generates fixed 81-frame windows stepping by 76 until the output reaches the driving video length, then trims the overshoot (228 driving frames = exactly 3 windows). Without one, it does a single window. The first 81 frames are identical to the stock single-chunk graph.

Requires a ComfyUI build with SCAIL-2 core support (comfy_extras.nodes_scail, added in PR #14373, merged 2026-06-09). On older builds this one node is skipped with a warning and the rest of the pack still loads. Models: Comfy-Org/SCAIL-2.

Nodes

| Node | Inputs | Outputs | |---|---|---| | Grok Vision Analyze (Image) | IMAGE tensor + prompt | response, usage | | Grok Vision Analyze (URL) | image URL string + prompt | response, usage | | Load Image From URL | URL string | IMAGE, MASK, url | | SCAIL-2 Infinity (auto window) | model, vae, conditioning, pose video + masks, reference image, sampler settings | images, latent, total_frames |

Author

@diobrando0 on HuggingFace · @workordie on Civitai · DanielBartolic on GitHub.

License

MIT — see LICENSE.