MiniMax H3 LongForge v5.0 — FL2VA & REF2VA

Follow & support

AI Bob Public · AI DEV LAB

If LongForge helps you, support future development through the Tip button on my Civitai profile or an optional USDT donation:

• BNB Smart Chain: 0xdCf9846A1B0F2fe32E19dC9CFdDC5194CC13D108

• TON: UQCeukjBulZw9gni1jv8NLRgtmgHDQ6mUiS2qf3716XobIxA

• TRON / 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, appearance, camera movement and sound continuity.

What’s new since v4.7

• Expanded RefMod integration: advanced extraction, masks, loaders, axis controls, text conditioning, curves and bundle tools are built into LongForge. No separate RefMod node pack required.

• Universal RefMod Studio: create image, video, audio or mixed bundles; inspect previews and token costs; repack existing references; save configurations and curve presets.

• New visual layouts: INDEPENDENT keeps separate references, STACK joins them in time, and MERGE blends multiple examples into one compressed reference. Optional refinement improves compressed-latent reconstruction.

• More reference controls: select individual bundle members, adjust visual/audio strengths separately, repeat references, switch between A/B files and apply frame or sampling-step curves. Bundles support up to 256 references per file, subject to memory and token limits.

• Stronger Native Cache protection: cache requires 07 · Turbo LoRA to be bypassed. Enabling that loader disables caching regardless of filename, strength or step count. Conflicting caches and active RefMod sampling-step curves also disable Native Cache.

• Stability fixes throughout: safer file saving and configuration updates, corrected mask alignment and reference numbering, improved media validation, cancellation handling and Studio settings persistence. Both workflows include updated Russian and English instructions.

How saved RefMods work

RefMods store reusable VAE latents in .safetensors files. Creation requires the matching H3 VAEs; it does not train H3 weights or a LoRA.

One file can contain multiple separate references. Packing photos into a single file does not automatically merge them.

• FULL: retains the encoded latent at the selected resolution.

• COMPRESSED: reduces token cost at the expense of detail.

• MERGE: creates a blended compressed reference and may soften different poses or backgrounds.

Saved latents are reused during generation without repeating source VAE encoding. Visual references are still decoded for text/vision conditioning. FULL is not lossless image storage, and reference strength does not guarantee identity or voice preservation.

RefMods require REF2VA, start disabled and support per-scene INHERIT / OFF / CUSTOM selection.

Main features

• FL2VA / REF2VA, first/last-frame inputs, ordinary media references and saved RefMods.

• Persistent scene projects, previews, regeneration, variations and Seam Lab for inspecting joins and comparing takes.

• OFF / HARD / SOFT audio continuity, optional 100 / 200 ms soft transitions, manual or automatic overlap.

• Optional Adaptive Native Cache, HQ latent upscale/refinement and local Gemma 4 Prompt Assistant.

• Audio references that follow the film timeline, complete MP4 export, continuous audio decoding, optional click repair and diagnostic WAVs.

Reference fidelity, speedups and scene continuity depend on the models and generation settings.

Included files

• ComfyUI-H3-LongForge-NodePack/

• H3_LongForge_FL2VA_REF2VA_v5.0.json

• RefMod_Studio_v5.0.json

Instructions are built into both workflows. Model weights are downloaded separately.

Quick start

1. Close ComfyUI and replace the previous LongForge folder under ComfyUI/custom_nodes/. Keep your saved projects.

2. Install the node pack’s requirements.txt using ComfyUI’s Python environment. Restart ComfyUI and refresh the browser with Ctrl+F5.

3. Open the main workflow, select the H3 model, text encoder and both VAEs, then match Pipeline to the model.

4. Enter a unique project name, write scene prompts and use GENERATE NEXT → Run.

5. Use EXPORT FILM → Run to assemble the active saved scenes.

For saved references, open RefMod Studio, add your sources, select CREATE, choose the source kind, mode and layout, then enable Run Studio action → Run.

Files default to ComfyUI/models/refmods/longforge/. Enable RefMods in the main workflow’s RefMod Library, select your file and use Check mapping to confirm its prompt labels.

To preview a saved reference, use INSPECT → stored → preview_index. The black 64 × 64 image after CREATE is a placeholder.

Requirements & links

Requires ComfyUI with native MiniMax H3 and V3 node support, matching H3 models/VAEs and working FFmpeg. The supplied main workflow also uses KJNodes and compatible SageAttention.

• MiniMax H3 models

• ComfyUI-KJNodes

• Optional Gemma 4 text encoders

• Optional H3 3D latent upscaler

RefMod tools are adapted from ComfyUI-MiniMaxH3Mod by Luisa / Luisacaotica. Third-party credits and licenses are included in the node pack.