Interested in AI video generation and open source https://xitter.cf/t.co/rItpNvosal
Joined July 2025
- Tweets109
- Following55
- Followers61
- Likes30
Having way too much fun with h3 swap lora. Swap: Characters, background, and style. All in one workflow.
More #h3 character swap lora fun. I threw this one at claude, so it could explain later how it was put together. It misses a few details like the hair and bowl but its overall a very impressive replacement of 2 characters. Sorry no sound, the original clip was missing it.
I just tried #h3 char swap lora/wf from @Alissonerdx , I was really blown away how good it works. And the example workflows intuitive. I swapped out the 2nd pass upscale with my own. A little complicated to get derope working in this pipeline so some blur but otherwise great.
I'm happy to see my upscale repo was helpful! Biggest thanks for @MiniMax_AI providing the open source community with such a capable model and supporting it's community 👍
More reference control. Fewer denoising steps. More places to run H3.
The latest community contributions span 8-step reference-conditioned video + audio, 4-step distillation, sparse attention, and reference-token compression - alongside serving and workflow integrations across GPU servers, TPUs, and Apple Silicon.
What matters is what this enables: more affordable experimentation, finer creative control, and more choices about where and how video generation runs. Open models become more useful when the community can improve the methods, adapt the runtime, and build workflows around them.
Credit to @haoailab @yisun0618 @lanfeng491270 @guocheng_qian @Alissonerdx @ddalcu @tazmannner @HaochengXiUCB @huggingface @Jhuaplin, and the PTD, DMAD, SparkH3, TorchTPU and FlashML teams - including Zhengming Yu and zechengtang on GitHub.
11 additions to the H3 ecosystem index. Thank you to everyone building, testing, and sharing.
github.com/MiniMax-AI/awesom…
tazmannner retweeted
More reference control. Fewer denoising steps. More places to run H3.
The latest community contributions span 8-step reference-conditioned video + audio, 4-step distillation, sparse attention, and reference-token compression - alongside serving and workflow integrations across GPU servers, TPUs, and Apple Silicon.
What matters is what this enables: more affordable experimentation, finer creative control, and more choices about where and how video generation runs. Open models become more useful when the community can improve the methods, adapt the runtime, and build workflows around them.
Credit to @haoailab @yisun0618 @lanfeng491270 @guocheng_qian @Alissonerdx @ddalcu @tazmannner @HaochengXiUCB @huggingface @Jhuaplin, and the PTD, DMAD, SparkH3, TorchTPU and FlashML teams - including Zhengming Yu and zechengtang on GitHub.
11 additions to the H3 ecosystem index. Thank you to everyone building, testing, and sharing.
github.com/MiniMax-AI/awesom…
Open weights. Shared progress. MiniMax H3 is moving fast.
We built H3 for video generation with native stereo audio and multimodal reference control. The open-source community is making that capability faster, more accessible, and easier to build on.
Recent highlights:
• FastH3 — FastVideo, Nuva Lab and NVIDIA: 4-step distillation, now running on DGX Spark and Apple Silicon.
• Sol-H3 — NVIDIA’s SANA team: 15 seconds of 768p video + audio in 6.6 seconds on 8×B300, in the team’s warm-inference benchmark.*
• VDN — Haocheng Xi and the OpenVDN team: rethinking attention for faster H3 inference, with weights, training and inference code released.
• PDD — NVIDIA’s distillation method, brought to H3 by Alibaba PAI as 8-step Acc-LoRAs, now supported in ComfyUI.
• LightX2V — 4- and 8-step Turbo LoRAs, with workflows for text, image and reference-conditioned video + audio.
Behind every release are people training, optimizing, quantizing, testing and sharing. Special thanks to:
@haoailab @nuvalab @NVIDIAAI @xieenze_jr @HaochengXiUCB @ArashVahdat @julberner @LightX2V @ComfyUI
And to the individual contributors pushing the work forward:
@haozhangml @cxlcl1 @lawrence_cjs @yitongli165665 @haopengl33 @songhan_mit @shanasaimoe
Thank you for building with H3 and helping make it faster, more accessible, and more useful for the community. Powerful models go further when we build together. Keep pushing H3. Excited to see what comes next. 🚀
Explore the ecosystem: github.com/MiniMax-AI/awesom…
I've been finding myself generating outside of comfy a lot lately. Just sending instructions to claude via remote control and then claude using comfyui for me. A lot easier but also more lazy...
Another small update to the #H3 anime upscaler, splitting the high/low sigmas on the first pass and using a custom node to make it easier. Speeds up the upscale pipeline 5-10%.
github.com/tazmannner/ComfyU…
For #H3, I trained an upscaler model specifically on anime and a custom workflow/node to optimize it further. A 0.5mp ->2mp upscale takes around x4 time LESS to generate at no quality loss compared to 2mp native. It also works with my other anime efficiency nodes:
github.com/tazmannner/ComfyU…
#H3: Generate 0.7mp / Switch to golden boy style -> Context Loop node for 3 prompts in 1 video -> face detailer + animate on 2s + 2mp high res fix. Added an additional scene at the end as well.
Revisiting an old #h3 prompt but with new workflows. This had everything render at 0.7mp, then context loop for 3 prompts stitched, then cut the clips and do 2nd pass improving with nodes and resplice. Some blood warning:
I made a hires fix node for minimax h3 on top of LBH’s upscaler. render at 1mp, take it to 1080p in about half the time of native 2mp, and the take doesn’t change, same acting and audio. 492s -> 237s at 2mp:
github.com/tazmannner/ComfyU…
A bit issue I've had up until now is that when you upscale on low resolution, the high noise has not had enough time to form the proper shapes. So you will sacrifice speed for quality. Around 0.7mp you can get close enough to full quality, but there isn't much to gain from a 0.7-1 or 1.5mp upscale. So I focused on creating the efficiency on the high end, going from 0.7 or 1mp to 2mp faster. Understanding that the first pass should not be toyed with too much.