RiverField

RiverField ComfyUI nodes

Small, self-contained nodes used by the RiverField studio. None of them modify another pack's files; the two planners only build the JSON documents that existing packs already expect as widgets.

Nodes

RiverField H3 Planner

Builds the timeline_data document of the MiniMax H3 Material Planner (MiniMaxH3TimelinePlanner). Wire its timeline_data output into the planner node.

| Input | What it is | | --- | --- | | segment_windows | JSON written by a timeline panel: windows with start/end in 24 fps panel frames, a prompt per window, optional media assignment | | global_prompt | scene, characters, style | | generation_seconds | segment length when there are no windows | | start_seconds, fps | timeline origin and frame rate | | segment_count, segment_prompts | form-driven alternative to the panel | | image1..image9, video1..video3, audio1..audio3 | media slots, empty slots are ignored | | audio_mode, video_audio, two_stage, upscaler, high_steps, advanced_timeline_json | planner options |

Output: timeline_data (string). Windows are snapped to H3-legal 5 + 17n frame lengths.

RiverField LTX Planner

Builds the three strings the LTX 2.5 Director (LTXDirectorCS25) expects, and turns an optional reference sheet into the still guide clip its IC-LoRA motion track reads.

| Input | What it is | | --- | --- | | segment_windows | same panel JSON as above (24 fps panel frames) | | global_prompt | scene, characters, style | | duration | total length when there are no windows | | fps | render frame rate (25 fps matches the LTX 2.5 templates) | | reference_sheet | optional image in ComfyUI's input folder: characters, props and sets side by side | | reference_lora | IC-LoRA applied when a sheet is given (Ingredients by default) |

Outputs: timeline_data, local_prompts, segment_lengths, duration_frames, duration_seconds, frame_rate, start_frame, end_frame, ic_lora_name.

Lengths are converted by seconds and snapped to the LTX 8n + 1 grid. A reference sheet is held on every frame with ffmpeg -loop 1 (the Director's motion track only reads video files) and the chosen LoRA name is emitted for the guide node.

Gemma 4 attention fix (startup)

Not a node: on import, ROCm builds only, it pins comfy.text_encoders.gemma4 to split attention. On ROCm 10.1 / gfx1151 the SDPA kernel traps on the Gemma 4 encoder of LTX-2.5 (hipErrorProfilerNotInitialized); the global --use-split-cross-attention workaround makes H3 video quadratic in memory, so the fix is scoped to the encoder. CUDA and CPU builds are left untouched.

RiverField Upscale Video

Chunked ESRGAN (or any spandrel) upscale for long clips. One frame at a time, one chunk file every chunk_seconds, concatenated with ffmpeg stream copy, then the source audio is muxed with -c:a copy. No full-clip decode in memory: a 2 h video stays at a flat RAM footprint. keep_chunks keeps the per-chunk mp4s for a manual resume after a crash.

Install

ComfyUI Manager: search for RiverField, or install riverfield-comfy-nodes.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ai-joe-git/riverfield-comfy-nodes

Restart ComfyUI. The nodes appear under the RiverField category.

Requirements

• ComfyUI (tested on 0.43.0).
• For the planners: the pack that owns the target widget — MiniMaxH3TimelinePlanner (Comfy-Org MiniMax H3) or LTXDirectorCS25 (CGlide's LTX-2.5-Director). The planners do not import them; they only produce strings.
• ffmpeg on the PATH for the LTX reference sheet clip.

License

MIT. See LICENSE.