A practical set of third-party tools has formed around the MiniMax H3 (Hailuo) video model: hosting provider fal.ai opened the H3 Max Multi Angle endpoint with frame-by-frame camera control, a 6-step turbo merge for GPUs with 16 GB of VRAM was released on Hugging Face, and the ComfyUI-H3-Continuum sampler allows assembling long videos without seams. This is not a new model release, but an engineering layer around it that makes controlled video generation cheaper and accessible locally.

image
image

What happened

The author of the 'Neuronaut' channel published the 21st link roundup on the MiniMax H3 (Hailuo) video model. On fal.ai, the H3 Max Multi Angle (image-to-video) endpoint is open, which provides frame-by-frame 3D camera control: azimuth, elevation, and distance with a 'frozen' scene. A 75% discount is in effect until September 14: 768p costs $0.02/s instead of $0.08/s, a 6-second 1080p video example costs $0.04/s, and under the promo rate, the price drops to $0.0125/s for 480p. On Hugging Face, the 6-step HardGravy2 turbo merge for ComfyUI was published, combining the turbo LoRA from lightx2v and larryvrh with the photorealism LoRA JonXL; on an RTX A4000 16 GB card at 0.5 MP resolution, with the Euler sampler and 6 steps, 10 seconds of video are generated in approximately 330 seconds. For long videos, the ComfyUI-H3-Continuum sampler is available with end-to-end AV-latent continuity, chunked sampling, and session resumption: the repository has 83 stars, and the update is dated September 11. The roundup also describes a local 2.6B GGUF prompt enhancer (LFM2 architecture, approximately 2.7 billion parameters, 131K context), which runs in llama.cpp or Jan.ai, and ready-made Fun Control and SEED HUNTER workflows from Civitai are available.

Context

MiniMax H3 (Hailuo) is a video generation model from MiniMax, and the roundup does not describe a new release of the base model: all the novelty lies in third-party infrastructure and the community. fal.ai is a model hosting provider with an API and per-second billing, Hugging Face is a platform for model distribution, and ComfyUI is a node-based environment for building generative pipelines. The term 'turbo merge' means combining LoRAs—auxiliary models that add specific properties, such as speed or photorealism—into a single ready-made build, and the '6-step' format means completing generation in six sampler iterations, which is important for speed with limited VRAM. The number 'Part 21' in the title is an episode of a regular series: the channel has long been collecting releases around MiniMax H3, and the appearance of three functionally different tools in one episode is a marker that a mature engineering ecosystem has formed around the model.

Why this matters for the industry

The ecosystem of open tools around MiniMax H3 is growing faster than for many closed video models: within a week, frame-by-frame camera control from a hosting provider, a 6-step turbo merge for 16 GB GPUs, and a sampler with latent continuity for long videos appeared. For companies and developers, this is a reproducible local pipeline on a 16 GB card and a production API with per-second pricing: A/B variants of advertising scenes and prototypes can be run without a cloud budget, and the cost per second of generation drops by a factor. For API providers, camera control is becoming a competitive factor: if the pace is maintained, 6-step distillations, seamless scene continuation, and depth/pose control will become standard practice in open ComfyUI video pipelines, and for science, the first independent H3 benchmarks on quality versus number of steps and long video stability will appear.

Why this matters for users

While the 75% discount is in effect until September 14, you can try precise camera control directly on fal.ai: 768p generation costs $0.02/s instead of $0.08/s, and a 6-second 1080p example costs $0.04/s. An owner of a card with 16 GB of VRAM can run the HardGravy2 turbo merge in ComfyUI and get 10 seconds of video at 0.5 MP in approximately 330 seconds. Long videos without seams are assembled through ComfyUI-H3-Continuum, including resuming an interrupted session, and prompts can be rewritten by a local 2.6B model in Jan.ai without sending the text to the cloud. Ready-made Fun Control and SEED HUNTER workflows are on Civitai; according to the author of the roundup, all seven links have been checked and are working.

What is still unknown / limitations

No community tool—HardGravy2, ComfyUI-H3-Continuum, SEED HUNTER—has published experiments or A/B quality comparisons, so the pipeline works without measured quality. 'Frame-by-frame' camera control is presented as an API feature without methodology: it is unclear whether this is true frame-by-frame control or interpolation between keyframes, there is not a single quantitative metric for camera pose accuracy, and it is unclear whether the feature is available outside fal.ai. The timing of ~330 s for 10 s of video is the only reproducible quantitative point without a comparison of what 6-step distillation loses relative to longer generation. HardGravy2 is assembled from third-party LoRAs without documented provenance and licenses, so reproduction may differ, and the legal cleanliness of the components is not confirmed. The 75% discount expires on September 14, and the post-promo price is not disclosed in the materials.

Sources

Author

Look at AI, editorial team