(Simple Workflow) MiniMax H3 — 10Eros + Semantic Bridge + Ultimate Upscale (8GB T2V/I2V) - v1.0 Showcase
Semantic Bridge — an ~11 MB conditioning adapter (not a second video model) that blends a learned semantic residual into H3 text cond. Goal is tighter prompt structure (spatial relations, stated actions, materials), not a style LoRA and not extra VRAM.
# MiniMax H3 — 10Eros + Semantic Bridge + Ultimate Upscale (8GB T2V)
Type: Workflow
Base: MiniMax H3 (FL2VA / text-conditioned)
Checkpoint: 10Eros TURBO-hybrid INT8 skip-edges
Example: SFW golden-hour causeway, native stereo audio
Download the JSON, drop the files below, restart ComfyUI, queue. You get the original clip and the upscaled clip.
More graphs: [artrealmai.com](https://artrealmai.com)
---
## What this is
A two-stage MiniMax H3 text-to-video graph for consumer 8GB cards:
1. Stage 1 — 10Eros TURBO-hybrid generates a short AV clip at 0.3–0.4 MP in 6 steps euler / simple). Turbo is already baked into 10Eros. Do not add LightX2V / ModelTC / other speed LoRAs.
2. Semantic Bridge — an ~11 MB conditioning adapter (not a second video model) that blends a learned semantic residual into H3 text cond. Goal is tighter prompt structure (spatial relations, stated actions, materials), not a style LoRA and not extra VRAM.
3. Stage 2 — [MMH3 Ultimate Upscale](https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale) with the official H3 3D latent upscaler, 8GB tiles, denoise 0.18.
H3 is CFG-distilled. This graph uses BasicGuider — no negative prompt, no CFG. Put exclusions in the positive prompt.
SenseNova is not required at runtime. Semantic Bridge was distilled from a teacher; inference is H3 + the 11 MB adapter only.
---
## Custom nodes (2 extra packs)
Everything else is ComfyUI 0.33+ native MiniMax H3.
1. [Comfyui-MMH3-UltimateUpscale](https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale) → custom_nodes/
2. [MiniMax H3 Semantic Bridge](https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge) — extract MiniMax_H3_Semantic_Bridge_v1.0.zip into custom_nodes/ (not in ComfyUI-Manager)
Restart ComfyUI after both are in place.
---
## Files
diffusion_models
[10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot_skip_edges.safetensors](https://huggingface.co/cicalooo/10Eros-Max-h3-int8-convrot/resolve/main/10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot_skip_edges.safetensors)
text_encoders
[qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors)
vae
[minimax_h3_video_vae_fp16.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors)
[minimax_h3_audio_vae_fp32.safetensors](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors)
latent_upscale_models
[minimax_h3_latent_upscaler_3d_fp16.safetensors](https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/resolve/main/minimax_h3_latent_upscaler_3d_fp16.safetensors)
semantic_bridge models/semantic_bridge/)
[MiniMaxH3_SemanticBridge_v1.safetensors](https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge/resolve/main/MiniMaxH3_SemanticBridge_v1.safetensors)
CLIPLoader type = minimax.
```
📂 ComfyUI/
├── 📂 custom_nodes/
│ ├── 📂 MiniMax_H3_Semantic_Bridge/
│ └── 📂 Comfyui-MMH3-UltimateUpscale/
└── 📂 models/
├── 📂 diffusion_models/ 10Eros skip-edges INT8
├── 📂 text_encoders/ qwen3vl NVFP4
├── 📂 vae/ video fp16 + audio fp32
├── 📂 latent_upscale_models/ H3 3D upscaler fp16
└── 📂 semantic_bridge/ MiniMaxH3_SemanticBridge_v1.safetensors
```
---
## 8GB defaults (graph as shipped)
| | Stage 1 | Stage 2 |
|---|---|---|
| Size | 9:16 @ 0.4 MP | 9:16 @ 0.9 MP (1280×736) |
| Sampler | euler / simple / 6 | euler / simple / 6 |
| Denoise | 1.0 | 0.18 |
| Attention | Comfy Kitchen | same model |
| Semantic Bridge | alpha 0.10, per_token | same prompt, matching W/H |
| Tiles | — | 512×384 (set ≥ upscale size to disable) |
| Temporal | — | chunk 1020 = off for a 5s clip |
Launch: --lowvram --force-fp16 --use-ck-attention
Load Image is bypassed = text-to-video. Enable it and wire first_frame only for I2V.
Outputs: video/MiniMax_H3_Original and video/MiniMax_H3_UltimateUpscale
---
## Prompt format
H3 wants a structured prompt, not a tag salad. The bundled example is a single continuous golden-hour shot on a wooden beach causeway (sundress, ocean, no on-screen text), with overall_soundscape and non_diegetic_music so the native audio track has something to do.
Keep Semantic Bridge alpha at 0.10 first. 0.15 is stronger (that is what the adapter author used for A/B clips). It can help stated constraints; it can also over-steer a shot. It is experimental on 10Eros (the adapter was distilled on native H3 FL2VA conditioning, which this graph still uses).
Do not use this adapter on Ref2VA.
---
## Tips
- Do not add turbo / speed LoRAs on 10Eros.
- Do not connect a negative prompt. H3 ignores it and you pay for a second pass.
- If VRAM dies on an 8s clip, drop temporal chunk_length to 85 and overlap 17.
- Extra style LoRAs: add ComfyUI's built-in Load LoRA (Model Only) after UNET. This share does not ship unpublished optional-LoRA nodes.
- Higher upscale denoise than 0.18 tends to make strange artifacts.
---
## Credits
- MiniMax H3 — MiniMax / Comfy-Org packages
- 10Eros TURBO-hybrid — TenStrip / cicalooo INT8 skip-edges
- Semantic Bridge adapter + nodes — [speach1sdef178](https://huggingface.co/speach1sdef178/MiniMax-H3-Semantic-Bridge)
- MMH3 Ultimate Upscale — [bbaudio-2025](https://github.com/bbaudio-2025/Comfyui-MMH3-UltimateUpscale)
- Graph — ArtRealmAI / [artrealmai.com](https://artrealmai.com)
---
## License (read this)
Free to use for personal and commercial generation. Videos you create with this graph are yours.
You may not sell, rebundle, paywall, or republish this workflow JSON (or a trivial copy) without permission from ArtRealmAI. Ask at [artrealmai.com](https://artrealmai.com).
Upstream MiniMax H3 Community License still applies to the H3 weights and to Semantic Bridge as an H3 model derivative. 10Eros and the upscaler have their own pages.