LTX 2.5 ReDetail Pass

The name on the label is a lie in the useful direction. "MiniMaxReDetailPass" sounds like something you bolt onto a MiniMax H3 render, but the weights it loads are LTX 2.5's - the "MiniMax" is a leftover from when this pack only drove MiniMax H3. And "Pass" hides the bigger surprise: this is not a node you drag onto a canvas.

It's the engine behind the ReDetail answer on the resolution pill in ComfyUI-Continuity (the repo formerly known as ComfyUI-MiniMax-Creator). Set the pill to ReDetail, and the render loop writes this node into the graph for you, after every pass is generated and every seam has taken what it needed. It re-renders each finished pass on the reel at twice the canvas through Lightricks' pixel spatial upscaler IC-LoRA. It's dev-only and filed under Continuity/internal for a reason: you turn it on, you don't wire it up.

A repaint, not a polish

Here's the honest framing, straight from the pack's source: ReDetail invents the fine detail as it goes - it's a repaint. Bambushu's original ReDetail (where this graph and its measurements come from) reports faces gaining freckles they never had and a motocross jersey coming back with different markings: stable across frames, but not the original's.

That's exactly right for AI-generated footage, which is soft and has no real detail to recover. It's wrong wherever a face has to stay the same person or a logo the same logo. If you need pixel-faithful, the pill's other answer - "two passes" - resolves what the first pass already drew. Different promise, different button.

How it works

The reel goes in, a re-detailed reel comes out. The node takes the reel (a list of passes), finds every part marked a pass, and re-renders each in play order inside a single call. One node means one set of loaders - the LTX transformer (~21.5 GB), text encoder (~15.4 GB), and both VAEs load once, at the end of a render that may have spent an hour in another family's weights.

The IC-LoRA goes straight onto the transformer, and its metadata answers GetICLoRAParameters with the downscale factor. That factor is the whole trick: the finished pass is encoded as a guide at strength 1.0, and because the factor is 2, the guide lands at exactly its own size - half the target. Prompts are empty; describing a shot that already exists would be a second opinion.

• Each part's first frame is pinned at 0.7 before the guide goes on, so a chunk doesn't drift from where it starts.
• Eight steps at CFG 1.0 on the distilled LTX 2.5 schedule - fixed, not your sampler row; the upscaler has one curve.
• The soundtrack survives untouched: the model gets the audio as a reference because it denoises a packed AV latent, but whatever it generates is thrown away - the rewritten pass points at the original file.
• LTX 2.5 silently drops the tail past 8n+1 frames; the pass is padded up with its last frame and the padding dropped after decode.

The inputs that matter

If you somehow do have this on a canvas (it's dev-only, so mostly you won't), the fields you'd actually set:

• ic_lora - the x2 pixel spatial upscaler IC-LoRA from models/loras. No LoRA, no pass - it errors and tells you to pick one. This file is what makes it a re-detail rather than a re-render.
• width / height - the finished size, twice what was sampled, on the /64 grid the guide's dilation needs (defaults 1536×2688).
• reel - the passes to re-render; model, clip, vae, and audio_vae are LTX 2.5's own slots.

The single output is a reel with the re-detailed passes swapped in at the new geometry, still pointing at the original audio.

Install

Same as any Continuity node - the pack isn't split up:

cd ComfyUI/custom_nodes
git clone https://github.com/roadmaus/ComfyUI-Continuity

Restart ComfyUI, nothing to pip install. ReDetail needs LTX 2.5's four files (transformer, text encoder, video VAE, audio VAE) in their usual models/ folders, plus one more: the LTX-2.5-22b-IC-LoRA-Pixel-Spatial-Upscaler dropped in models/loras.

Where people get burned

VRAM is the tax. Bambushu measured 243 frames at 768×1408: 2× took 17 minutes at 80.5 GB, and the rule of thumb is frames × megapixels under ~150 at 24 GB, ~350 at 48 GB. This is the last upscale option a small card should enable.

• Long renders dying with HostBuffer.read_file_slice CUDA OOM are a known ComfyUI dynamic-VRAM issue - start ComfyUI with --disable-dynamic-vram.
• If a pass isn't at the canvas the node was built for, it errors loudly rather than quietly re-snapping: you loaded a workflow built against a different piece.
• Feathered seams. Across a hard cut each part inventing its detail independently is exactly right. Across a feathered seam - two passes that continue one another - the join can show. The author says so in the code, not the marketing.

If you render LTX through Continuity and have the headroom, ReDetail is worth a try - LTX's internal half-res render is precisely the soft-footage case this repaint was built for. Just don't expect the same person's face to come back identical.