RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2098331716111843330?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV16DYL6FEsh/
# LightX2V — Four-Step Reference Configuration
## Purpose and preparation
Resolve the exact model and node versions, inspect task/input agreement and choose authorized images before queueing. Treat saved controls as a starting point, not accepted output evidence.
Image 114 reaches reference slot 0; task AUTO. Active Ref2V LoRA strength 1; sigma shift video 12/audio 3. No bypassed generation branch; two MarkdownNotes are notes, not extra projects.
BasicScheduler explicitly stores simple, 4 steps, denoise1, with Euler selection. This is a genuine saved four-step setting, not a measured speedup, total-NFE result or quality acceptance. Different model variants and sigma shifts must not be collapsed into one identical configuration.
Selected dimensions are 896×512, length 124 with no external size/time override. Noise 9527 has fixed control. Sampler OUTPUT reaches video/audio decoders, then CreateVideo 24fps/8-bit and one SaveVideo with auto format/codec. No source-audio passthrough is connected.
At24fps, 124 output frames would be nominally about5.167 seconds, not a measured duration. All enabled soundtracks come from decoded generated audio; source audio inputs are absent and conditioner mux_audio is unused. Prompt soundscapes are instructions, not audible proof. Inspect actual task resolution, images, timing, speech and endpoints in a small authorized-media test.
## Files and selected assets
- [Workflow JSON]()
- [Selected cover]()
- `minimax_h3_video_vae_int8_convrot.safetensors`
- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`
- `minimax_h3_audio_vae_fp32.safetensors`
- `minimax_h3_ref2va_int8_convrot.safetensors`
- `minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors`
## Version and rights
New copies retain the original configuration except recorded historical preview cleanup. No task-mode mismatch is silently repaired. This is descriptive local copy, not a published upgrade.
No inference, visual/audio acceptance, hardware timing or license verification was performed. Covers are selected same-stem WebP assets, not accepted output samples. Weights and external input media are not bundled. Check workflow, nodes, models, LoRAs, images and cover permissions separately; no blanket commercial license or speed winner is asserted.
[Source-provided tutorial](https://www.youtube.com/watch?v=g3QmFCf2dF4) · [Source workflow entry](https://www.runninghub.cn/post/2098326332844363777/?inviteCode=rh-v1111)
Links not live-verified. The workflow is supplied through AI-KSK.
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/None?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV16DYL6FEsh/
# LightX2V — Four-Step Reference Configuration
## Purpose and preparation
Resolve the exact model and node versions, inspect task/input agreement and choose authorized images before queueing. Treat saved controls as a starting point, not accepted output evidence.
Image 114 reaches reference slot 0; task AUTO. Active Ref2V LoRA strength 1; sigma shift video 12/audio 3. No bypassed generation branch; two MarkdownNotes are notes, not extra projects.
BasicScheduler explicitly stores simple, 4 steps, denoise1, with Euler selection. This is a genuine saved four-step setting, not a measured speedup, total-NFE result or quality acceptance. Different model variants and sigma shifts must not be collapsed into one identical configuration.
Selected dimensions are 896×512, length 124 with no external size/time override. Noise 9527 has fixed control. Sampler OUTPUT reaches video/audio decoders, then CreateVideo 24fps/8-bit and one SaveVideo with auto format/codec. No source-audio passthrough is connected.
At24fps, 124 output frames would be nominally about5.167 seconds, not a measured duration. All enabled soundtracks come from decoded generated audio; source audio inputs are absent and conditioner mux_audio is unused. Prompt soundscapes are instructions, not audible proof. Inspect actual task resolution, images, timing, speech and endpoints in a small authorized-media test.
## Files and selected assets
- [Workflow JSON]()
- [Selected cover]()
- `minimax_h3_video_vae_int8_convrot.safetensors`
- `qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`
- `minimax_h3_audio_vae_fp32.safetensors`
- `minimax_h3_ref2va_int8_convrot.safetensors`
- `minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors`
## Version and rights
New copies retain the original configuration except recorded historical preview cleanup. No task-mode mismatch is silently repaired. This is descriptive local copy, not a published upgrade.
No inference, visual/audio acceptance, hardware timing or license verification was performed. Covers are selected same-stem WebP assets, not accepted output samples. Weights and external input media are not bundled. Check workflow, nodes, models, LoRAs, images and cover permissions separately; no blanket commercial license or speed winner is asserted.
[Source-provided tutorial](https://www.youtube.com/watch?v=g3QmFCf2dF4) · [Source workflow entry](https://www.runninghub.cn/post/2098326332844363777/?inviteCode=rh-v1111)
Links not live-verified. The workflow is supplied through AI-KSK.