ComfyUI News & Open-Source AI Releases

Dials turns prompt pairs into 14 KB sliders for MiniMax H3 in ComfyUI: motion from near-still to 2.7x, eight evil styles, Pika H3 Flex support and word dials.

KohakuBlueleaf's MOSS-Image-heal retrains MOSS-Image-Base so its residual stream stops spiking: fp16 no longer runs near the limit, and ComfyUI gets bf16, fp16 and int8 files.

Viggle Remake is an open MiniMax H3 LoRA pair for recasting or editing existing footage: paint one to four frames, keep the soundtrack, run it in ComfyUI down to 12 GB.

minimax-h3-community-license

H3 Relay Hires renders a MiniMax H3 take at 0.5 MP in ComfyUI, then finishes at 1920x1088 with an anime upscaler and one DMAD step: 119 s instead of 503 s on an RTX 5090.

MIT

NitroFusion2 distills Z-Image-Turbo to one and two steps, ships native ComfyUI workflows and GGUF files from 2.0 GB, plus int8 files about 1.7x faster per step.

Krea 3 edits images from up to eight references at native 2K, finishes under 8 seconds and costs $0.06. It runs in ComfyUI through the Krea 3 Image node.

API-Only

vivo open-sources BluePixel-Turbo, a FLUX.2 Klein editor with a DCAE-F64 VAE that edits at native 4K in 4 steps, with a community ComfyUI node pack.

Idol Turbo distils Speridlabs Idol from 50 guided steps to five guidance-free ones, 100 network evaluations down to 5, with two phase LoRAs that drop into the ComfyUI Idol repack.

Apache-2.0

ComfyUI SPEED adds KSampler-shaped nodes so progressive-resolution sampling runs on MOSS Image and Krea 2 as well as MiniMax H3, with per-model calibration and new timings.

PrunaAI's P-H3-Image renders 8 MP images in 8 steps with no CFG and edits with up to 8 references. ComfyUI added native support the day it launched.

minimax-h3-community-license-agreement

Anima RGBA pairs a 4-channel Qwen-Image VAE with an Anima LoRA so the model renders transparent PNGs with no custom nodes, at alpha MAE 0.0027 on a held-out set.

H3 Continuity Studio is a ComfyUI node pack that lays MiniMax H3 shots on a timeline and feeds five-frame visual memory into the next clip, so rooms and outfits survive a cut.

LightX2V's Lightning v2 RL swaps the 67 MB adapter for a full Qwen-Image 2.1 transformer distilled to 4 steps and refined with reinforcement learning.

Qwen Research License

Plenio 1.5 for ComfyUI sings your own lyrics as written, writes style template values into the Song Brief, and lifts creativity to 200.

ZenTrainer adds LoRA training to ComfyUI for MOSS Image, Krea 2, FLUX 3, Qwen-Image 2.1, Z-Image, Anima, Fibo 2.0, SDXL and MiniMax H3, with live previews and AMD support.

MIT

H3 Ripple is a MiniMax H3 LoRA that takes one or two frames you edited by hand and propagates the change across the shot, with ComfyUI workflows for external or local Qwen edits.

Tessera is a ComfyUI node pack that refines MiniMax H3 clips tile by tile into 2K, 4K and 8K, with per-tile conditioning and streamed encode and decode.

Ostris published Primal VAE, two ~80M parameter autoencoders that downsample only 2x, leave the 8x patchify to the DiT, and carry no scaling or shift factor.

Apache-2.0

SnapRef v1a keeps an Anima character from a single reference image by caching its attention keys and values, at 8.9 s per image. ComfyUI nodes and FP8 weights included.

The LightX2V team retrained RelayVSR for native 2x video super-resolution, with 45 FPS on an H100, flicker suppression and two checkpoints.