RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2098775143422537729?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV1vFYz6HECU/
# Sigma two-pass — 4K-target post-processing
## Purpose and preparation
Resolve the matching custom-node/runtime versions and assets. Load authorized media and inspect active inputs, actual schedules, effective dimensions and every save destination before a small test.
Only image 139 is active. Image256 remains wired but bypassed; images 257–260 are bypassed/disconnected. The generation selector requests 16:9,0.5MP,multiple32, overriding cached dimensions. Duration 5 requests 124 frames arithmetically, not a measured five-second output.
The active Ref2V LoRA is named 8step, approximately strength 1. BasicScheduler stores simple/10/denoise1, followed by H3SigmaRefiner(extra_steps 1,start 0.7,end 0,cosine), then SplitSigmas at 6. Because refinement precedes the split,6+4 actual steps are not established. The model sigma shift is video 12/audio 3; Euler, BasicGuider and fixed-control noise are shared.
First DENOISED_OUTPUT is separated. Video enters a learned FP16 latent upscaler configured multiplier 2/align 32/CUDA/chunking; audio bypasses that upscaler and is concatenated into the second latent. Second OUTPUT feeds video/audio decoding. This is not proof of unchanged audio, exact raster or quality improvement. Native generated audio, not source soundtrack or conditioner mux_audio, reaches assembly.
Decoded images pass through SeedVR2, linked target2160, then a resize request3840×2160, Lanczos, keep_proportion=resize, divisible_by2, CPU. Audio bypasses spatial processing. These are post-decoding targets, not native generation dimensions or measured results. Cached resize diagnostics are not render evidence.
Interpolation requests 2x,motion auto,threshold0.24,keep_audio on. This is temporal, not spatial scaling. For generation presets,24→48fps is configured intent, not measured cadence/frame count/duration. Combined VIDEO and explicit images/audio are connected; precedence requires implementation review. VIDEO output feeds SaveVideo with auto format/codec.
## Files and selected assets
- [Workflow JSON]()
- [Selected cover]()
- `qwen3vl_32b_minimax_h3_int8_convrot.safetensors`
- `minimax_h3_video_vae_fp16.safetensors`
- `minimax_h3_audio_vae_fp32.safetensors`
- `minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors`
- `minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors`
- `seedvr2_ema_3b_fp8_e4m3fn.safetensors`
- `minimax_h3_latent_upscaler_3d_fp16.safetensors`
- `taeh3.safetensors`
## Version, evidence and rights
This is a local descriptive draft, not a published upgrade. Files01–08 are byte-identical to originals; only the standalone ninth file's historical video-preview field is cleared in its new copy. Originals remain unchanged.
No model inference, visual/audio acceptance, wall-clock benchmark or license verification was performed. Covers are selected assets, not accepted render evidence. Weights and external input media are not bundled. Check workflow, nodes, models, LoRAs, source media and cover rights separately; no blanket commercial license is asserted.
[Source-provided tutorial](https://www.youtube.com/watch?v=R3W2ab1_TzY) · [Source-provided workflow entry](https://www.runninghub.cn/post/2098400619433381889/?inviteCode=rh-v1111)
Links are not live-verified. The workflow is supplied through AI-KSK.