RunningHub workflow page: https://www.runninghub.ai/post/2101698407000969217?inviteCode=rh-v1111

Bilibili tutorial video: https://www.bilibili.com/video/BV1hEhh6JEvE/

# Two-Pass Native-AV Refinement

## Purpose and preparation

Use the matching custom nodes and authorized reference media. Inspect bypass modes, emitted conditioning, dimensions, schedules and save destinations before a small test.

Three reference-image loaders are active; first/last-frame, source-video and all source-audio loading routes are bypassed. AUTO/native conditioning is configured. The linked prompt overrides cached conditioner text, with extra sharpness/photorealism wording for refinement. Verify sparse reference-slot/Picture ordinal handling in the installed implementation.

Coarse dimensions come from 4:3 / 1 MP / multiple 32. A learned latent upscaler configured scale_by 1.5 supplies refined dimensions. Old 736×416,1344×768 and inactive target-size fields are not the effective resolution specification. The 10-duration control reaches an alignment expression giving 243 requested frames arithmetically; nominal 243/24=10.125 seconds is not a measured video duration.

Turbo and ExposureSlider LoRAs are bypassed. Active coarse selections are Combat 0.7, aesthetic 0.5, FrameRush 0.5 and Motion Repair 0.7. Refinement uses Combat 0.49, aesthetic 0.5, FrameRush 0.5, Motion Repair 0.2 and LMS 0.3. These labels do not prove de-oiling, face restoration or identity preservation; no dedicated face-crop/stitch route is present.

Parity-plan controls 9/4/5 provide connected coarse/refine schedules. The base sampler's old 8-step sigma output is unused. The first sampler's DENOISED_OUTPUT goes through learned upscale and reconciliation into refinement. Tail, time-bias, STG and restart are disabled despite dormant nonzero knobs. No actual NFE, speed or memory claim is made.

Main output decodes refined video and generated audio, with VHS 24 fps/H.265/yuv420p10le/CRF22. Two additional stage decode/create/save branches are active, so inspect all save destinations. Stage2 decodes the same refined latent again; its wrapper settings are not the main VHS settings. Generated audio is not source-audio passthrough. Four stale producer-side link references remain documented, not repaired.

## Files

- [Workflow JSON]()

- [Selected cover]()

## Selected active/auxiliary assets

- `Motion_Repair.safetensors`

- `minimax_h3_lms_v1.0_r64.safetensors`

- `10Eros_Max_h3_TURBO-hybrid_beta5_int8_SinoX.safetensors`

- `minimax_h3_video_vae_fp16.safetensors`

- `minimax_h3_audio_vae_fp32.safetensors`

- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`

- `ig_tiktok_aesthetic_h3_lora_v1_500.safetensors`

- `H3_Combat_V2.safetensors`

- `FrameRush-Minimax-V2_c2-st1500.safetensors`

- `minimax_h3_latent_upscaler_3d_fp16.safetensors`

## Version and evidence boundaries

Only historical preview metadata is cleared in new copies; original files and other graph data are preserved, including documented stale references. This is descriptive local copy, not a published update.

No new inference, visual/audio review, timing or license acceptance is included. Selected covers are not verified generated results. Model weights and external input media are not bundled. Check workflow, node, model, LoRA, image, sound and cover rights separately; there is no blanket commercial license.

[Source-provided tutorial](https://www.youtube.com/watch?v=KjHrHot5un8), not live-verified. The workflow is supplied through AI-KSK.