MiniMax H3 vs Wan 2.2 · AI video head-to-head
MiniMax H3 vs Wan 2.2: who wins?
75 blind duels · same prompt, names hiddenSame prompt, side by side
Not enough votes on this prompt yet.
How these clips were made
The exact settings behind the clips at the top. Open models run on our own GPU in ComfyUI: download the workflow and render the same clip yourself.
- diffusion modelFP8 scaledminimax_h3_fl2va_pruned_fp8_scaled.safetensors
- LoRABF16strength 1minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors
- text encoderNVFP4 (AWQ)qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
- VAEFP16minimax_h3_video_vae_fp16.safetensors
- VAEFP32minimax_h3_audio_vae_fp32.safetensors
- Sampler
- res_multistep
- Scheduler
- simple
- Steps
- 8
- CFG
- 1
- Seed
- 1040068153
- Size
- 1024×576
- Frames
- 124 @ 24 fps
- Render time
- 5 m 44 s
- GPU
- NVIDIA GeForce RTX 5060 Ti · 16 GB
- Engine
- ComfyUI 0.39.0 · PyTorch 2.11.0+cu128
- diffusion modelFP16wan2.2_ti2v_5B_fp16.safetensors
- text encoderFP8 e4m3umt5_xxl_fp8_e4m3fn_scaled.safetensors
- VAEwan2.2_vae.safetensors
- Sampler
- uni_pc
- Scheduler
- simple
- Steps
- 20
- CFG
- 5
- Shift
- 8
- Seed
- 964652333
- Size
- 1024×576
- Frames
- 121 @ 24 fps
- Render time
- 5 m 34 s
- GPU
- NVIDIA GeForce RTX 5060 Ti · 16 GB
- Engine
- ComfyUI 0.39.0 · PyTorch 2.11.0+cu128
Negative prompt: blurry, low quality, static frame, distorted, deformed, watermark, text, subtitles
ComfyUI workflow (.json)Full prompt our open models got
The short prompt, expanded into a shot description (look, camera, timeline, sound) before rendering. Closed models expand prompts themselves, so they get the short one.
Shot on anamorphic lens with warm golden-green grading, soft shafts of sunlight piercing humid haze, cinematic and tense. Two women stand face to face in a dense, lush jungle clearing, ornate armor plates and flowing layered dresses in worn bronze and deep teal. Ferns, vines and broad leaves crowd the background, moss-covered roots line the ground, and sun rays filter through the canopy, catching drifting pollen and mist. Both hold a fighting stance with feet planted and fists slowly tightening at their sides. Camera: Begins at a medium two-shot, then a lateral tracking move circles slowly around the pair, shifting to a low-angle close-up on the closer fighter as the tension builds, before pulling back to a wider frame. Timeline: [0s-2s] Both women hold still, shoulders rising with breath, fists clenching; light shifts subtly through the leaves. [2s-4s] The camera orbits in, dust and leaves stirring from the ground as the first woman steps forward into a sharper stance. [4s-6s] Fast, controlled movement: both women lunge and sidestep, armor plates and fabric swinging with the motion while the camera follows at close range. [6s-8s] The camera rises and pulls back to a wide shot as both settle back into their stances, sunlight flaring through the canopy. Realistic anatomy, consistent costumes and faces throughout, smooth natural motion.
| Spec | MiniMax H3 | Wan 2.2 |
|---|---|---|
| Provider | MiniMax · open weights | Alibaba · open weights |
| Elo | 1055 | 975 |
| Win rate (all duels) | 65% | 40% |
| Max length | 5s | 5s |
| Resolution | 576p | 576p |
| Native audio | Yes | No |
| Price | Free (our queue) | Free (our queue) |
| Where to try | FrameFight free queue | FrameFight free queue |
| Speed | ⚡⚡ | ⚡ |
Voters prefer MiniMax H3 in 63% of 75 blind duels.
Pick MiniMax H3 if you need free, audio, readable text and don’t mind waiting in the free queue. Pick Wan 2.2 if free, open weights, landscapes matter more — and it runs free in our queue.
Watch out: Wan 2.2 is most often flagged for 720p; MiniMax H3 for 576p on our gpu.
Try your prompt in both — freeRankings come from votes only. Some links are ads — how we make money.