RunningHub workflow page: https://www.runninghub.ai/post/2101698905921818625?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV1hEhh6JEvE/
# SelfLift Source-Audio Conditioning
## 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.
Two images are connected; the saved prompt assigns character and environment/composition roles. Those descriptions do not verify actual image content. ResolutionSelector 3:4 / 1.5 MP / multiple32 overrides old landscape dimensions. Native reference-audio and reference-video ports are unconnected.
TrimAudioDuration stores start_index 14.63 and duration 20.46, not a start/end pair. Check units and available media with the matching implementation. Trimmed source audio is VAE-encoded, receives a zero noise mask, and is concatenated with the video part of the reference template before sampling.
Reported audio length passes through an INT output before the frame formula. The formula aligns to 17n+5, then adds 24. Illustrative integer inputs 5/10/20 yield 148/267/505, not observed outputs. Do not treat fractional source duration, padding and output synchronization as already verified.
The four-step LoRA is bypassed. External scheduling stores simple/6/denoise1 with Euler. SelfLift stores CFG 1, transition 3, lowres_scale 0.5, rho 0, w_min 0.5, w_max 1 and the learned 3D latent upscaler; model_hires is unconnected. This is not a measured six-evaluation benchmark or proof of exact intermediate dimensions. Negative conditioning is ZeroOut.
The sampler AV result is split and only video is decoded. VHS receives trimmed source audio directly and stores 24fps/H.264/CRF12, trim_to_audio=false. There is no explicit postdecode trimming node. This establishes routing, not bit-perfect audio, correct lip-sync or equal track duration. Three stale producer-side references are documented; imported connections require review.
## Files
- [Workflow JSON]()
- [Selected cover]()
## Selected active/auxiliary assets
- `10Eros_Max_h3_TURBO-hybrid_beta5_int8_SinoX.safetensors`
- `qwen3vl_32b_minimax_h3_int8_convrot.safetensors`
- `minimax_h3_video_vae_fp16.safetensors`
- `minimax_h3_audio_vae_fp32.safetensors`
- `minimax_h3_latent_upscaler_3d_fp16.safetensors`
- `taeh3.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.
Description
RunningHub workflow page: https://www.runninghub.ai/post/2101698905921818625?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV1hEhh6JEvE/
# SelfLift Source-Audio Conditioning
## 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.
Two images are connected; the saved prompt assigns character and environment/composition roles. Those descriptions do not verify actual image content. ResolutionSelector 3:4 / 1.5 MP / multiple32 overrides old landscape dimensions. Native reference-audio and reference-video ports are unconnected.
TrimAudioDuration stores start_index 14.63 and duration 20.46, not a start/end pair. Check units and available media with the matching implementation. Trimmed source audio is VAE-encoded, receives a zero noise mask, and is concatenated with the video part of the reference template before sampling.
Reported audio length passes through an INT output before the frame formula. The formula aligns to 17n+5, then adds 24. Illustrative integer inputs 5/10/20 yield 148/267/505, not observed outputs. Do not treat fractional source duration, padding and output synchronization as already verified.
The four-step LoRA is bypassed. External scheduling stores simple/6/denoise1 with Euler. SelfLift stores CFG 1, transition 3, lowres_scale 0.5, rho 0, w_min 0.5, w_max 1 and the learned 3D latent upscaler; model_hires is unconnected. This is not a measured six-evaluation benchmark or proof of exact intermediate dimensions. Negative conditioning is ZeroOut.
The sampler AV result is split and only video is decoded. VHS receives trimmed source audio directly and stores 24fps/H.264/CRF12, trim_to_audio=false. There is no explicit postdecode trimming node. This establishes routing, not bit-perfect audio, correct lip-sync or equal track duration. Three stale producer-side references are documented; imported connections require review.
## Files
- [Workflow JSON]()
- [Selected cover]()
## Selected active/auxiliary assets
- `10Eros_Max_h3_TURBO-hybrid_beta5_int8_SinoX.safetensors`
- `qwen3vl_32b_minimax_h3_int8_convrot.safetensors`
- `minimax_h3_video_vae_fp16.safetensors`
- `minimax_h3_audio_vae_fp32.safetensors`
- `minimax_h3_latent_upscaler_3d_fp16.safetensors`
- `taeh3.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.