embriotapir/LTX-2.5-MSR-DualStage-Workflow

LTX-2.5 MSR — Dual-Stage Upscale Workflow (ComfyUI)

A tidied-up, ready-to-use ComfyUI workflow for LTX-2.5 Multi-Subject Reference (MSR) video generation, with an optional 2x latent-upscale refinement stage you can toggle on/off with one click.

Built on top of ComfyUI-LTX2.5-MSR — this repo just packages it into a clean, organized workflow with subgraphs, shared controls, and a stage-2 bypass toggle so it's easier to pick up and use.

Features

• Multi-subject reference conditioning — up to 5 reference images (pic1–pic4 + background) blended into one generated video via learned MSR slot embeddings.
• Dual-stage generation — a fast low-res base pass, followed by an optional 2x latent-upscale + low-noise refinement pass for extra detail.
• One-click Stage 2 toggle — the whole upscale stage is bundled into a single group with a Fast Groups Bypasser switch (via rgthree-comfy). Turn it off for quick previews, on for final renders.
• Single source of truth for prompts/settings — FPS, DURATION, and the two prompt fields (global_prompt reference descriptions + local_prompts action/narrative) are each set once and shared across both stages using Set/Get nodes, so nothing gets out of sync between stage 1 and stage 2.

Requirements

Custom node packs

Install these via ComfyUI-Manager or manually into custom_nodes/:

• ComfyUI-LTX2.5-MSR — MSR LoRA loader + multi-reference guide nodes
• ComfyUI-PromptRelay — prompt relay / temporal segment text encoding
• ComfyUI-KJNodes — constants + Set/Get utility nodes
• rgthree-comfy — Fast Groups Bypasser (stage 2 toggle)
• ComfyUI-VideoHelperSuite — video combine/export

Models

Place these in the standard ComfyUI model folders:

Type	Folder	Notes
LTX-2.5 diffusion model (UNET)	diffusion_models/	distilled variant recommended for speed
LTX-2.5 text encoder (CLIP)	text_encoders/	Gemma-based, per official LTX-2.5 release
LTX-2.5 video VAE	vae/
LTX-2.5 audio VAE	vae/
MSR LoRA	loras/	must contain reference_slot_embedding weights — a normal LoRA will be rejected
LTX-2.5 latent spatial upscaler (x2)	latent_upscale_models/	only needed if you use Stage 2

How to use

1. Open the workflow in ComfyUI.
2. Upload your reference images into the PIC 1–PIC 4 / Background Load Image nodes (any of pic2–pic4/background can be left empty — the MSR guide skips missing optional references).
3. Edit the two shared prompt boxes:
   • Global Prompt — static descriptions of each reference image (what's in Image 1, 2, 3...).
   • Local Prompts — the action/narrative describing how they combine in the final video.
4. Set FPS and DURATION (in seconds) — video length is computed automatically.
5. Queue the prompt.
6. Want a quick preview without the upscale pass? Open the Fast Groups Bypasser node and flip the "Stage 2 Upscale" toggle off — only the fast base pass will render. Flip it back on for the full dual-stage render.

VRAM notes

This workflow was built and tested on a 24GB card (RTX 3090). A few tips if you hit CUDA out-of-memory, especially during Stage 2:

• Launch ComfyUI with --disable-pinned-memory if you get OOM errors that look inconsistent with the reported VRAM usage — pinned-memory allocation can be unreliable on some systems.
• Lower DURATION (fewer frames = much lower memory for the upscale pass, since attention/FFN activations scale with frame count).
• Enable use_tiled_encode on the MSR Multi-Reference Guide nodes and set a reasonable tile_overlap if you're still tight on memory.
• Bypass Stage 2 entirely for iterating on prompts/composition, and only enable it for your final render.

Credits

• MSR LoRA training, IC-LoRA loader, and multi-reference guide nodes: ComfyUI-LTX2.5-MSR by LiconStudio
• Sample reference images / inspiration for this setup: thanks to LiconStudio's LTX-2.3-Multiple-Subject-Reference for providing samples to work from
• Prompt Relay temporal encoding: ComfyUI-PromptRelay
• Set/Get constant nodes: ComfyUI-KJNodes
• Stage 2 bypass toggle: rgthree-comfy
• Video export: ComfyUI-VideoHelperSuite
• Base model: LTX-2.5 by Lightricks

License

This repo only contains the ComfyUI workflow file (a JSON graph description) — no model weights are redistributed here. Model, LoRA, and text-encoder files referenced by the workflow are subject to their own respective licenses; download them from their original sources linked above.