← sources

Introducing H3 Relay Hires node: 0.5 → 2 MP anime upscaler, ~4x faster than native, with a full 20-step pass 1 (no turbo LoRA needed for the take)

source: reddit (reddit:StableDiffusion)published 2026-10-28 04:15ingested 2026-10-29 02:40

retained text

Repo: [**https://github.com/tazmannner/ComfyUI-H3-RelayHires**](https://github.com/tazmannner/ComfyUI-H3-RelayHires) (MIT, 8 example workflows)

Also I started posting lots of progress on more anime stuff I'm doing in open source on twitter:

[https://x.com/tazmannner](https://x.com/tazmannner)

I worked with claude to make this tutorial video and write up. This is still a work in progress and appreciate any constructive feedback.

I got tired of waiting 8+ minutes for every 1920 H3 clip, so I made a hires method for it: a 2 MP clip about 4x faster than a native render. It doesn’t rely on running the whole render on a low-step LoRA, which in my experience brings trade-offs like poorer motion and distorted visuals or colors. You render the take at 0.5 MP like a preview. If you like it, the same take gets finished at 1920x1088, and the audio stays exactly as you heard it.

**On a 5090, for a 5 s clip at 1920x1088, end to end:**

|Render type|time|
|:-|:-|
|native 1920 render|\~500 s|
|1 MP + latent upscale + refine|\~240 s|
|**Relay Hires**|**\~118 s**|
|**Relay Hires + Fast Tail Sampler**|**\~107 s**|

**How it works.** It's all one node:

1. **Pass 1 at 0.5 MP.** This is where H3 decides the layout, motion, acting and audio, and it's where most of the time goes. You approve the take here.
2. **AnimeSR upscale to 1920 (\~2 s).** It's a tiny upscaler I trained from scratch, and it's included in the repo. H3 draws \~2 px ink lines at *every* resolution, so normal upscalers give you lines twice too thick and mushy small faces. AnimeSR is trained to undo exactly that.
3. **One H3 turbo step at 1920** (DMAD LoRA, from sigma 0.60). This redraws the detail at full size. The pass-1 audio is pinned, so the soundtrack is identical.

The H3 step only redraws detail, so you get the take you previewed, not a new one. A latent upscale + refine can change faces and poses; this doesn't. Smeared fast frames and tiny faces in wide shots also come out clean. Still lines barely shimmer: 0.13 frame to frame, vs 0.10 native and 0.22 for the 1 MP route.

**Honest limits:**

* **Line weight:** ink lines are a bit bolder than a native 1920 render.
* **Very fast action:** if the 0.5 MP take itself melts (a sword clash), the relay gives you motion blur there, not crisp poses. Example workflow 2 fixes those frames first with a de-rope pass.
* **Anime only:** it's built for 2D anime. I haven't tried realistic stuff.

**Bonus: pass 1 \~20% faster.** The pack also has an **H3 Fast Tail Sampler**. H3 decides the take in its first 12 of 20 steps, so those run exactly like euler, and the last 8 become 4 second-order steps. Same take, 58 → 47 s for pass 1. If you'd rather not install it, the README shows how to wire the same thing with stock nodes.

# How to use it

**You need:**

* **MiniMax H3** in ComfyUI. (Duh)
* The **DMAD 4-step turbo LoRA for H3** (Kijai's MiniMax-H3-experimental repo on Hugging Face).
* **taeh3** (Kijai's MiniMax-H3-TAE) and a loader for it. The examples use the one from [ComfyUI-H3-FastDerope](https://github.com/tazmannner/ComfyUI-H3-FastDerope).
* **VideoHelperSuite** and **comfyui\_essentials**. (not a hard requirement but reccomended)

**Steps:**

1. `git clone` [`https://github.com/tazmannner/ComfyUI-H3-RelayHires`](https://github.com/tazmannner/ComfyUI-H3-RelayHires) into `custom_nodes`. The AnimeSR model is already inside.
2. Open **example workflow 1** (`1_relay_hires.json`). It's the simplest one and doesn't need any of my other packs besides FastDerope's taeh3 loader.
3. In the blue **SETTINGS** group: write your prompt (one box, used everywhere), and leave the size at **0.5 MP** and the hires scale at **2.0**.
4. Queue. You get the 0.5 MP preview and the 1920 final.
5. Optional (See below) but you can just integrate the workflow into your existing workflow quite easily. And it works for both FLV2A and REF since the magic is done mostly in the 2nd pass.

**Tip:** keep the seed fixed. Mute the Relay Hires node and the final save (Ctrl+M) while you try prompts and seeds, then unmute when you like the preview. ComfyUI reuses the finished pass 1, so you only pay for the relay.

**Adding it to your own workflow.** Put the **H3 Relay Hires** node after your pass 1 and connect:

* **model:** your pass-1 model (same LoRAs and shift) plus the DMAD LoRA at 0.8.
* **positive:** your pass-1 conditioning.
* **latent:** the pass-1 sampler output. The audio comes from here.
* **images:** your decoded pass-1 frames.
* **vae:** taeh3.

Then LTXV Separate AV Latent → VAE Decode. Keep the defaults: AnimeSR, sigma 0.60, 1 step, window 0.

**The other example workflows** each add one stage before the relay:

* **2:** de-rope, for fast-motion smears (144 s).
* **3:** \+ FaceDetail, for small faces (224 s).
* **4:** \+ Anime Cadence, for 2s/3s holds (228 s).
* **5-8:** the same four with the Fast Tail Sampler.

Those need my other packs, linked in the README (3-4 also use two small helper nodes from ResCascade, to put the faces and holds onto the 1920 frames). If there is interest I may make posts on these later, they all have their uses depending on the situation.

**I've also used it for:**

* **Character swap / ref2va:** works unchanged. A 12-shot, 22 s character-swap recast at 1920 took only \~3 min more than the same run at 0.5 MP.
* **15 s clips and multi-scene chains:** relayed part by part.

# Bonus: my other anime packs for H3

All free, and they work with the relay (example workflows 2-4 use them):

* [**Fast De-rope**](https://github.com/tazmannner/ComfyUI-H3-FastDerope)**:** fixes H3's fast-motion smearing and melting. A motion gate finds the fast moments and redraws only those, then splices them back into your render. On one 10 s clip with a single fast kick, the fix cost +12 s instead of +202 s. The de-rope itself is matlowai's [MAINodes](https://github.com/matlowai/ComfyUI-MAINodes); this pack makes it cheaper and cleaner so credit to him for his great work.
* [**Anime Cadence**](https://github.com/tazmannner/ComfyUI-H3-AnimeCadence)**:** makes H3 move like hand-drawn anime. It puts drawings on 2s/3s where H3 slips into a new picture every frame, and keeps the sharpest frame of each held drawing. A character-only mode holds the character while the background and camera keep moving smoothly.

Everything here is on a 5090 with 96 GB RAM. I haven't tested smaller cards, so I'd love to hear how it runs on yours.

Built with a lot of help from Claude (Claude Code), which did a ton of the testing and measuring.

document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xhecjk/introducing_h3_relay_hires_node_05_2_mp_anime