H3-Pika RealVSR RTX VFX — Update 20261108 - For Patreon Supporters (Patreon)
Videos
-
11月8日.mp4 -
779661853.mp4
Downloads
Content
Hey everyone.
If you watched the RelayVSR video, you saw what native video super-resolution does to soft AI output.
https://www.patreon.com/aifuturetech/posts/relayvsr-video-174673261
This post is the full pipeline from the video: generate with MiniMax H3, regenerate up to 2K with Open-H3-Regenerate, upscale with RelayVSR, then finish with the RTX pass (frame generation and DLSS5 enhancement). One workflow, attached to this post: H3-Pika_RealVSR-RTX_VFX.json.
----------------------------------------
WHAT TO DOWNLOAD (MODEL FILES)
----------------------------------------
New for this video (RelayVSR, Apache-2.0):
models/diffusion_models/
relayvsr_x2_lightx2v_int8_convrot.safetensors
models/vae/
relayvsr_flash_vae_x2_fp32.safetensors
Download both from: https://huggingface.co/Comfy-Org/RelayVSR
(the original upstream is https://huggingface.co/kopper/RelayVSR)
H3 side — you will have most of this from the earlier H3 videos:
models/diffusion_models/
minimax-h3/Minimax-h3_Singularity_ref2va_v2_int8.safetensors
models/text_encoders/
minimax-h3/qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors
models/vae/
minimax-h3/minimax_h3_video_vae_fp16.safetensors
minimax-h3/minimax_h3_audio_vae_fp32.safetensors
models/loras/
minimax-h3/h3-realism-people-t2v-i2v-r2v(r34l1sm).safetensors
minimax-h3/minimax_h3_lms_v1.0_r64.safetensors
pika_h3_flex/pika_h3_flex_turbo_8step_hd_ref2va.safetensors
Open-H3-Regenerate-2K/open_h3_regenerate_2k_lora.safetensors
Open-H3-Regenerate-2K/sharp_curve_branches.safetensors
Loaded by the Open-H3-Regenerate nodes (same folder setup as that release):
minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors
fast_h3_vae_noncausal_24l.safetensors
Plus RMBG-2.0 for the character-sheet background removal.
Software side: use a current ComfyUI build. RelayVSR support merged on November 5, 2026 (PR 17511), and it needs the comfy kitchen attention backend, so an older install will not see the model at all.
----------------------------------------
QUICK START
----------------------------------------
1. Load H3-Pika_RealVSR-RTX_VFX.json into ComfyUI (current build).
2. Check the loaders: H3 UNET, CLIP, the two H3 VAEs, and the RelayVSR subgraph (x2 model + flash VAE). It's a ComfyUI subgraph — double-click into it if you want to see the conditioning and decode nodes inside.
3. Load your character sheet images into the two LoadImage nodes. Background removal and resizing are wired in, so raw sheets are fine.
4. Write your prompt in the subject_definitions format below.
5. Set the duration. The 10.0 float means seconds. The math node turns it into a frame count padded for H3's 17-frame chunking. Never edit the frame count directly.
6. Seeds ship fixed at 42. Change them only when you want a brand-new generation.
7. Queue. You get the base H3 render, the 2K regen, and a video-compare strip (input and output side by side) so you can judge the RelayVSR pass directly.
That's the whole loop.
----------------------------------------
WHAT TO INPUT (PROMPT FORMAT)
----------------------------------------
The prompt has two parts: the quality tag prefix and the subject_definitions block.
Part 1, the prefix (concatenated in front of everything):
r34l1sm , ultra realistic video.
This is the trigger the h3-realism LoRA listens for. It ships at strength 0.85, keep it.
Part 2, the subject block. One subject per character, each tied to a picture slot that matches your LoadImage order:
subject_definitions:
<Subject 1> is the ice sovereign whose appearance comes from <Picture 1>: a tall, pale woman in her late twenties with platinum-blonde hair in a long thick braid, a crown of blue ice crystals, ...
Rules that matter:
- Concrete details beat adjectives. "platinum-blonde hair in a long thick braid, a crown of blue ice crystals" gives both H3 and RelayVSR something to lock onto. "beautiful queen" gives them nothing.
- Every appearance element you describe is a detail the upscale pass can preserve. If you want it sharp in 2K, name it in the prompt.
- Two LoadImage nodes means up to two <Picture N> slots. Match the order.
----------------------------------------
KEY SETTINGS (AND WHY)
----------------------------------------
Sampling structure — 12 steps total, split at step 6. The first six steps generate at 1344x768 (16:9, factor 0.8 in the ResolutionSelector). The second six feed the Open-H3-Regenerate stage, which runs 3 steps on 2K (2560x1440) latents using the 3D-conv latent upscaler. Neither stage runs long, which is why the whole pipeline stays reasonable on time.
MiniMaxH3SigmaShift 12 / 3 — base gen and regen sigma shifts. Ships correct, leave it.
MiniMaxChunkFeedForward 2 chunks / 4096 tokens — the VRAM saver on the H3 model. Leave as shipped; lower it only if you hit out-of-memory.
BlockSparseAttention sol-attn 1.3, sparsity 0.2, thresholds 12288 / 256, exact_kv_and_rows — keeps long clips manageable. Ships correct, leave it.
RelayVSR subgraph — relayvsr_x2_lightx2v_int8_convrot + relayvsr_flash_vae_x2_fp32, with detail exposed at 1. The subgraph wraps RelayVSRConditioning and VSRDecode, and the exposed detail input is the same knob as VSRDecode's detail in the free upscale workflow: 0 collapses toward bicubic, values below 1 reduce the sharpening, above 1 pushes texture harder. Start at 1 and move in 0.1 steps.
RTXVideoFrameGeneration — frame rate multiplier 2, HIGH quality. Doubles the 24 fps output to 48 fps. If your GPU labors, drop quality to MEDIUM first, multiplier second.
DLSS5 — 1x (DLAA / native) with Enhance Images on. It cleans up noise and detail at native resolution instead of changing size, so it stacks cleanly on top of RelayVSR.
Everything else (euler, beta scheduler, ModelPatchTorchSettings, comfy kitchen attention backend) ships correct — leave as is.
----------------------------------------
TIPS
----------------------------------------
1. Judge RelayVSR on the video-compare output, never on the raw render. The compare strip puts input and output side by side on the same frame, and small sharpening differences read wrong when you flip between separate videos.
2. Tune one variable at a time with the seeds fixed. Both seeds ship fixed so a re-run with one changed setting is a true A/B.
3. Duration changes go through the seconds float only. H3 generates in 17-frame chunks and the math node pads the count so clips land on a chunk boundary. Editing frame count by hand breaks that and you get a seam.
4. If you only want RelayVSR on existing footage, run the free upscale-only workflow from the public post first. Same RelayVSR pair, no H3 generation, much faster feedback while you learn the detail knob.
5. The 2K regen is the heaviest stage. If VRAM is tight, drop the Open-H3-Regenerate target from 2K (2560x1440) to the 1080p tier before you cut the H3 base resolution — regen is where the spikes happen.
6. Faces drifting between chunks is an H3 chunking artifact, not an upscaling one. The lms LoRA (ships enabled at 0.8) helps; shorter duration helps more.
----------------------------------------
COMMON ISSUES
----------------------------------------
Out of memory during the 2K regen — lower the Open-H3-Regenerate target to the 1080p tier, or shorten the duration. Chunked feed-forward (2 / 4096) is already on, so the next lever is resolution.
Output looks over-sharp or plasticky after upscaling — drop the RelayVSR detail value below 1 (try 0.8), or lower the DLSS5 sharpness. The two knobs stack, so move one at a time.
RelayVSR model not recognized — ComfyUI is too old. You need a build from November 5, 2026 or later (PR 17511) with the comfy kitchen attention backend available. Update and restart.
Video-compare strip out of sync — an output node got a different frame rate. Keep every VideoCombine node at 24 fps and the strip stays aligned.
Error mentioning window attention or comfy kitchen — same cause as above. The RelayVSR model uses windowed causal attention over keyframes and the backend ships with recent ComfyUI. Update.
DLSS pass looks flat or dim — Enhance Images should be on, and HDR tone mapping lives in the free workflow's RTXTrueHDR node if you want to match the video look.
----------------------------------------
The workflow file is attached. Download the two RelayVSR files from the Hugging Face link above before you run anything.
Questions? Drop a comment with your GPU, VRAM, and the exact error — that trio turns most problems into a one-line fix.
Thanks for supporting the channel. You guys make these deep dives possible.
— Benji