Create pictures and videos in the browser — text to image, text to video, and free daily gens with ArtRealmAI Gen.
Generative AI: Reports · Model Drops · Prompts · Tutorials
A new one-paste installer runs Lightricks' open LTX-2.5 22B distilled model on Macs, Intel CPUs included. Patched GGUF loaders with Metal kernels stream the model from disk, so a 2-second clip with sound takes about 20 minutes on an 8 GB machine.
BlueRose Labs fine-tuned Kyutai's Pocket TTS 18M into Silencio-2, a CC-BY-4.0 English voice model that takes inline tags like [sigh], [laugh] and whisper cues, clones a voice from a short clip, and runs about 20x faster than real time on a plain CPU.
A new open toolkit runs MiniMax H3 video with audio on a 16 GB AMD RX 6900 XT under Windows through stable-diffusion.cpp and ROCm. Tuning cut one 480p clip from 62 minutes to about 3, and every fix is documented.
A community LoRA trained with inclusionAI's new Intrinsic Distribution Matching recipe gets MOSS-Image down to 2 steps with no CFG. That's 2 DiT passes instead of 100, with a free Space and a ComfyUI path you can use today.
A community Krea 2 fine-tune that places panels, characters and exact speech-bubble text inside boxes you draw in plain text just got INT8 builds. That's about 14 GB instead of 26 GB, and they drop straight into ComfyUI's stock Krea 2 template.
A SIGGRAPH Asia 2026 paper just shipped code and a 717 MB FLUX.1-dev LoRA that turns ordinary photos into full 360° HDR environment maps. It can even give you a different light probe for every spot in a room.
A new experimental launcher for native ComfyUI splits MiniMax H3's attention and MLP work into 1024-token chunks. On an 8 GB laptop RTX 5060, a 10-second 768×1344 clip with audio stopped running out of memory.
Jutsu Studio's new open-source MCP server lets Claude, Cursor or Codex price-shop Seedance, Kling, Veo, FLUX and friends across 43 providers (or your own ComfyUI), stop at a budget, and keep every result in a local library.
A new unofficial ComfyUI pack wraps vivo's ICLR 2026 Any-to-Bokeh model, so you can script a cinematic focus pull onto finished footage, keyframe by keyframe, without regenerating a single frame.
A new open-weights ComfyUI app hands LTX-2.5's Restore IC-LoRA one clear photo of the person, then shows you side by side how much that photo pulls a blurry home-video face back toward the real one.
A new MIT node pack finds garbled logos, odd faces and broken hands with SAM 3 and regenerates only those patches. It uses local Qwen Image Edit on a 10 GB GPU, or Nano Banana 2 and GPT Image 2 via partner nodes, and blends the result back pixel-perfect.
A new MIT node pack brings NVIDIA's SIGGRAPH 2026 ARDY motion model into ComfyUI. You click the ground to steer a live 3D body, record the take, and render it as driving frames for MiniMax H3 or as a GLB.
ComfyUI core now writes ProRes straight from Save Video, as 10-bit 4444 with alpha or 422 HQ in MOV or MKV, and adds a chroma-subsampling switch to Create Video. It was merged on November 1, so git users can pull it today.
A new Apache-2.0 DMD distill of NVIDIA's ChronoEdit-14B edits images in 4 network evaluations instead of 100, and it matches the 50-step model on GEdit-Bench. It went up on Hugging Face on November 1.
H3MD Director is a new open-source ComfyUI node pack that plans multi-shot MiniMax H3 videos, renders them segment by segment to 2+ minutes at 1080p, and re-renders only the bad segments. It went live on the Comfy Registry on October 31.
Lodestone's chroma-h3 fine-tunes the MiniMax H3 video model into a text-to-image stills model, with open weights, a rank-256 LoRA and a new 8-step DMD LoRA. A ComfyUI-format conversion is already up on Hugging Face.
Ian's MIT-licensed Ultra Elements packs 39 real-time spell effects for the browser, from fireballs to portals, and its plan shows how parallel AI agents built each element under a 3 ms frame budget.
WhiteTowerAI's Apache-2.0 OhMyGame lets an agent design, generate assets for, code, playtest and publish a game. Beta 5 adds bring-your-own coding agents, an MCP server, Godot, Unity and Unreal plugins, and interactive stories.
ZenTrainer 0.4.0 turns ComfyUI into a LoRA training studio with live previews, likeness scores and ten end-to-end tested models, from FLUX 3 and Krea 2 to MiniMax H3 video. It's free, MIT-licensed and runs on NVIDIA or AMD.
AInVFX's Paint a Plate is a free ComfyUI app for VFX artists: paint on the first frame to erase, add or relight, and the edit rides through the whole shot with FLUX 3 Klein, LTX 2.5 and SAM 3.1. It even exports in the plate's own format.
MiniMax H3 Max and Seedance 2.0 Fast take up to 9 images. Cite them as Picture or @Image tags. First 4 H3 images are included; each extra is 3¢. Seedance images are included in the per-second rate.
Text or image to video, draft or full, 5–20 seconds with audio. Grok and MiniMax are still on the rail.
Pin favorite gens to your public profile. Load a prompt back into Gen.
Stamp a logo on a folder of images or videos in the browser.
4 free stills a day on Flux.2 Klein 9B after you sign in (4 at 1K or 2 at 2K).
