MiniMax H3 LongForge v4.7 — FL2VA & REF2VA
Follow & support
If LongForge helps you, support future development through the Tip button on my Civitai profile or an optional USDT donation:
BNB Smart Chain:
0xdCf9846A1B0F2fe32E19dC9CFdDC5194CC13D108TON:
UQCeukjBulZw9gni1jv8NLRgtmgHDQ6mUiS2qf3716XobIxATRON / TRC20:
THLU3yu8VpoPqP4bMLCp86FRhoCweJ2Qz4
Please send only USDT using the matching network. Thank you!
Generate, continue, review and export long MiniMax H3 videos with audio in one editable ComfyUI workflow.
LongForge carries native video and audio latent context between scenes to help maintain motion, camera movement, appearance and sound continuity.
What’s new since v4.0
Saved RefMods: reusable image, video and audio references stored as
.safetensors, with file selection, aliases, strength and component controls. Built into LongForge; no separate RefMod node pack required.One universal RefMod Studio: create all three reference types in a single workflow using H3 VAEs. Supports FULL / COMPRESSED visual latents and optional foreground masks. No diffusion sampling or LoRA training is needed for creation.
Per-scene reference control: inherit the RefMod Library defaults, disable saved references for a scene, or select a custom set. Ordinary media and saved RefMods have independent reference policies.
Seam Lab: inspect scene joins with looped playback, frame stepping and audio diagnostics. Compare compatible saved takes and choose an alternative; later scenes then become pending.
Prompt Assistant fixes: improved text-generation compatibility, draft validation and support for larger reference maps.
Safer project controls: delayed server replies no longer overwrite a different project or a newly edited draft.
Cleaner release package: two workflows with readable Russian and English instructions built directly into the graphs.
Saved RefMods require REF2VA and start disabled. Their latents are reused, while visual references still require VAE decoding for H3 text conditioning during generation.
Main features
FL2VA / REF2VA with text, first/last-frame and media-reference inputs.
Persistent scene projects with previews, regeneration, variations and pending-scene reordering.
OFF / HARD / SOFT audio continuity, including optional 100 / 200 ms soft transitions.
Manual overlap or AUTO BALANCED / AUTO SAFE selection.
Optional Adaptive Native Cache, HQ latent upscale/refinement and local Gemma 4 Prompt Assistant.
Long-audio references that follow the film timeline.
Complete MP4 export with continuous audio decoding, optional click repair and diagnostic WAVs.
Optional workflow recovery metadata.
Keep Native Cache bypassed when using Turbo or another short-step acceleration setup. Reference fidelity, speedups and seamless joins depend on the model and generation settings.
Included files
ComfyUI-H3-LongForge-NodePack/H3_LongForge_FL2VA_REF2VA_v4.7.jsonRefMod_Studio_v4.7.json
All user instructions are inside the workflows. Models, VAEs, LoRAs and optional enhancement weights are downloaded separately.
Quick start
Close ComfyUI and replace the previous LongForge node-pack folder under
ComfyUI/custom_nodes/. Keep your saved projects.Restart ComfyUI, refresh the browser with Ctrl+F5 and open the main workflow.
Select the H3 model, text encoder, video VAE and audio VAE. Match Pipeline to the model.
Give your project a unique name, write scene prompts and use GENERATE NEXT → Run.
Choose EXPORT FILM → Run to assemble the active saved scenes.
For saved references, open RefMod Studio, select both H3 VAEs, add your source, choose IMAGES / VIDEO / AUDIO, enable Create & save and run. Then select the resulting file in the main workflow’s RefMod Library.
Files are saved under ComfyUI/models/refmods/longforge/. RefMods can be reused across projects.
The main workflow starts with FL2VA, 1344 × 768, 158 frames, 22-frame overlap and HARD audio continuity. RefMods and HQ Finish are OFF; video metadata is NONE.
Requirements & links
A compatible ComfyUI build with native MiniMax H3 support and the V3 node API, matching H3 models/VAEs, and usable FFmpeg are required. The supplied main graph also uses KJNodes and compatible SageAttention.