akhaliq/viggle-remake-turbo

Viggle Remake Turbo

Few-step rendering for Viggle Remake — recast people and edit scenes in existing videos, at 4 steps or even 2 steps instead of 6. Each file here is Remake's task LoRA fused with a proven few-step H3 distillation into one standard LoRA (exact rank-concatenation — no custom merge math, no approximation). Load a single file onto any stock Remake setup: ComfyUI, bf16 / int8 / GGUF. Video with the original soundtrack.

The lineup

File (both swap and edit modes)	Steps	Turbo half	Pick for
viggle_remake_turbo4x_*_comfyui.safetensors	4	lightx2v fl2v turbo v1.2 768p	Production — fl2va-native turbo, matches the transformer Remake's LoRA applies to
viggle_remake_turbo_*_comfyui.safetensors	4	SMD Ref2VA 4-step	Alt 4-step (v1, validated)
viggle_remake_turbo2d_*_comfyui.safetensors	2	pDMD 2-NFE (Apache 2.0)	Draft — check framing & motion fast, render finals at 4

Speed on one A100 80GB, compact profile (int8 transformer, nvfp4 text encoder), 832×480, 226 frames, seed 0, end-to-end including text encode and VAE/audio decodes — the step count is what scales:

variant	forwards	measured
Viggle Remake (reference, 6 steps)	6	~95 s on B200 (Viggle's measurement)
turbo 4-step (either turbo half)	4	140–170 s
turbo 2-step draft	2	80 s

2-step is draft-only: fine for iterating on framing and motion, but like every 2-step H3 distill it degrades audio and can drop scenes (pDMD's own guidance and the Pika Flex Turbo findings agree: finals at 4+ steps).

Quickstart (ComfyUI)

Same install as stock Remake (Viggle's COMFYUI.md), plus one file from here:

cd ComfyUI
hf download Comfy-Org/MiniMax-H3 \
  diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors \
  text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors \
  vae/minimax_h3_video_vae_int8_convrot.safetensors \
  vae/minimax_h3_audio_vae_fp32.safetensors --local-dir models
hf download Viggle/Viggle-Remake --include "comfyui/ComfyUI-Viggle-Remake/*" "examples/inputs/*" --local-dir ./Viggle-Remake
cp -r Viggle-Remake/comfyui/ComfyUI-Viggle-Remake custom_nodes/
cp Viggle-Remake/examples/inputs/swap-pineapple-mascot/source.mp4 input/remake_swap_source.mp4
cp Viggle-Remake/examples/inputs/swap-pineapple-mascot/ref1.png   input/remake_swap_ref1.png
cp Viggle-Remake/examples/inputs/swap-pineapple-mascot/ref2.png   input/remake_swap_ref2.png
hf download akhaliq/viggle-remake-turbo --include "*.safetensors" "workflows/*" --local-dir ./remake-turbo
cp remake-turbo/*.safetensors models/loras/

Queue a workflow from workflows/ (the stock example inputs/prompts are wired in). Every workflow keeps Viggle's graph intact and changes only the sigma grid and the turbo LoRA.

Two rules, inherited from the upstream docs:

• Load with LoraLoaderBypassModelOnly, never LoraLoaderModelOnly — Viggle measured the merging loader rounding much of a LoRA update away at bf16.
• Use the fl2va diffusion model file — Remake's LoRA applies to the base transformer, and every turbo half here composes onto the same target modules.

Sigma grids (already set in the workflows)

variant	ManualSigmas
turbo4x (lightx2v, shift 6)	1.0, 0.947368, 0.857143, 0.666667, 0.0
turbo (SMD, shift 12)	1.0, 0.997228, 0.990826, 0.96, 0.0
turbo2d (pDMD, shift 12)	1.0, 0.923077, 0.0

Euler sampler, no CFG — same as stock Remake. Pinned keyframes (RemakeKeyframeUntil 0.9) stay held for the whole denoise at these grids.

Samples

All rendered at the stock example settings (832×480, 226 frames, seed 0) — compare against Viggle's official 6-step clips.

Swap — pizza maker → pineapple mascot, turbo4x (lightx2v, 4 steps):

Swap — 2-step draft (pDMD):

Edit — wind turbines removed, turbo4x (lightx2v, 4 steps):

Edit — 2-step draft (pDMD):

Also in samples/: the v1 SMD renders (turbo_swap_pineapple.mp4 stacked, turbo_swap_pineapple_merged.mp4 from the merged file, turbo_edit_turbines.mp4) and swap_lx2v4_smdgrid.mp4 (lightx2v on the SMD sigma grid, for the curious).

Validation (build reports: build_report.json, build_report_v2.json)

• Exact composition, verified per layer. Every merged file is a rank-concatenation: B_fused @ A_fused == (B_remake @ A_remake) + (B_turbo @ A_turbo) per layer. A separate tensor check replayed all 210 modules in both modes: worst relative error 1.2e-3 (bf16 rounding level) — the merged file is the stacked adapters, as one file. Concatenation cannot interfere the way naive weight addition can; this is the same construction the Pika H3 Flex Turbo release uses, citing arXiv 2604.16826.
• Renders do the task. Every sample is far from its source clip and in the official-clip PSNR band (the swap example is chaotic: Viggle measures ~27 dB movement from bf16-level changes, and the official clip itself sits 11.4 dB from the source; all swap renders land 10.3–11.3 dB from official). The edit renders sit 16.8–18.0 dB from official against a 17.8 dB scale.
• Key compatibility verified before rendering: each turbo LoRA shares 200 target modules with the Remake LoRA (the rest are token-refiner / patch-proj layers unique to each).
• Merged-vs-stacked render PSNR 14.12 dB on the swap example: expected trajectory divergence on the documented-chaotic clip, not an error — the merged file's deltas are exact (above), and both renders pass the official/source checks independently.

Why stacking a task LoRA with a turbo LoRA works

This composition is the designed use case of acceleration LoRAs: LCM-LoRA (Eq. 5–6) defines generation as base + λ₁·task + λ₂·acceleration, and a task LoRA × step-distillation LoRA is the least-conflicting pair — one changes what is generated, the other changes how the ODE is traversed. The turbo halves run at strength 1.0, as their authors require (distillation deltas are calibrated, not style deltas). Method background: SMD, pDMD via Iwannapose's conversion, Compress-then-Merge and LoRA Soups on why concatenation beats merge-then-truncate.

Attribution and license

Community repack — all credit upstream:

• Viggle Remake © 2026 Viggle AI — Viggle/Viggle-Remake. If you build on it, credit "Powered by Viggle Remake" where users see it.
• Turbo halves: lightx2v/Minimax-h3-Turbo, zjwfufu/SMD, pdmd2026 via Iwannapose.
• Recipe and lineup format follow Pika-Labs/Pika-H3-Flex-Turbo-LoRA.
• MiniMax H3 — MiniMaxAI/MiniMax-H3, base files via Comfy-Org.

All weights here are Model Derivatives of MiniMax H3: the MiniMax H3 Community License applies, including display of "MiniMax H3" in commercial products, passing the license and NOTICE downstream, and written authorization above US$20M/year revenue. The identity in a Remake output comes from the frames you paint — only run it on people who have agreed to it, and label what you generate as AI-generated. Not affiliated with or endorsed by Viggle AI or MiniMax.