This workflow is a practical MiniMax H3 H25 hybrid reference-to-video setup for creators who want to combine several visual references, reference audio, long-form prompt direction, and a fast eight-step generation route in one organized ComfyUI graph. It is built around a tested parameter combination rather than a generic starter template. The main use case is a short cinematic sequence in which character identity, props, environment references, spoken audio, action timing, and camera instructions all need to remain coordinated across the same generation.

The active model route uses MiniMaxH3HybridLoader with `minimax_h3_fl2va_bf16.safetensors` and `minimax_h3_ref2va_bf16.safetensors`. The hybrid loader applies the H25 configuration through the connected `block_range_adaln` range, while `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` is loaded at strength 1.0 for the accelerated route. Text conditioning is handled by the Qwen3-VL 32B MiniMax H3 NVFP4 AWQ encoder. Separate MiniMax H3 video and audio VAEs are connected, allowing the final graph to preserve both generated motion and the audio side of the reference-driven workflow.

The conditioning section is designed for multi-reference production. Several image inputs feed the reference-to-video structure, while dedicated audio inputs support voice and sound references. A long structured prompt can define subjects, shot timing, dialogue, action beats, sound design, and music direction. The active graph uses a 16:9 widescreen resolution selector, an Euler sampler, an eight-step beta scheduler, MiniMax H3 sigma shifting, and SageAttention in automatic mode. The final output path decodes video and audio, creates the video at 24 fps, and saves the combined result. Image scaling nodes and the optional video-reference branch visible in the JSON are bypassed, so they are not presented as active features.

This setup is especially useful for fantasy action, dialogue scenes, product storytelling, character interaction, and any sequence where multiple visual references must play distinct roles. The included prompt demonstrates how to assign subjects, define exact time ranges, stage progressive action, and reserve clean audio space for dialogue. For fair testing, keep the same references and prompt when comparing it with the related parameter-test workflow.

Main features:

- MiniMax H3 FL2VA and Ref2VA hybrid model route
- H25 hybrid configuration through the connected loader
- LightX2V-style eight-step Turbo LoRA at strength 1.0
- Multi-image reference conditioning for characters, props, and scenes
- Reference audio inputs for dialogue and sound guidance
- Qwen3-VL 32B MiniMax H3 NVFP4 AWQ text encoder
- Separate MiniMax H3 video and audio VAE decoding
- 16:9 widescreen generation route
- Euler sampler with an eight-step beta scheduler
- MiniMax H3 sigma-shift controls
- Automatic SageAttention patching
- Structured long-prompt support with timed shots and dialogue
- Combined video and audio export at 24 fps
- Bypassed experimental branches kept available without being misrepresented

Suggested workflow:

Start by assigning every reference image a single clear role: character, environment, prop, creature, or style target. Use clean references with readable silhouettes and avoid giving two inputs the same semantic job. Add the audio references, then write the prompt with explicit subject labels and time ranges. Test a shorter, simpler action first to confirm identity and audio behavior before moving to a dense cinematic sequence.

Keep the eight-step route and connected H25 loader settings unchanged for the first comparison. If motion becomes unclear, simplify simultaneous actions and reduce rapid camera changes before changing model parameters. For dialogue, keep the speaking face visible and leave enough time for each line. Use the related RunningHub comparison page to evaluate alternate parameter combinations with the same source material.

RunningHub Workflow

Try the workflow online right now - no installation required.
Workflow: https://www.runninghub.ai/post/2088213692442173441?inviteCode=rh-v1111

Related parameter comparison:
https://www.runninghub.ai/post/2088221911972106242?inviteCode=rh-v1111

If the results meet your expectations, you can later deploy it locally for customization.

Fan Benefits: Register to get 1000 points + daily login 100 points - enjoy 4090 performance and 48 GB super power!

Bilibili Updates (Mainland China & Asia-Pacific)

If you're in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown.
Bilibili Video: https://www.bilibili.com/video/BV1dygg6EE6W/

Support Me on Ko-fi

If you find my content helpful and want to support future creations, you can buy me a coffee.
Every bit of support helps me keep creating.
Ko-fi: https://ko-fi.com/aiksk

Business Contact

For collaboration or inquiries, please contact aiksk95 on WeChat.

打开下方链接即可在线体验，无需安装。
工作流：https://www.runninghub.ai/post/2088213692442173441?inviteCode=rh-v1111

多参数对比测试：
https://www.runninghub.ai/post/2088221911972106242?inviteCode=rh-v1111

如果你觉得效果理想，也可以在本地进行自定义部署。

粉丝福利：注册即可领取 1000 积分，每日登录再领 100 积分，体验 4090 和 48 GB 大显存性能。

B站视频（中国大陆及亚太地区）

如果你在中国大陆或亚太地区，可以通过下面的视频查看工作流的实测效果与创作思路。
B站视频：https://www.bilibili.com/video/BV1dygg6EE6W/

我会在夸克网盘持续更新模型资源：
https://pan.quark.cn/s/07bdc81784ce