H3-Class video generation on a single RTX 5090: 15 seconds in 15, at 448 by 256. Comfy-Org's comfystreamerh3 node fits an 80 GB H3 video pipeline into 32 GB of consumer VRAM. The speed is real. The resolution is the catch. 10 October 20264 min read792 words What happened. Comfy-Org published comfystreamerh3, a ComfyUI node that runs H3-class video generation on a single NVIDIA RTX 5090, producing 15 seconds of 448 x 256 video in 15 seconds or less. The workflow stacks six hardware-level optimisations onto FastVideo's FastH3 V2 checkpoint to fit an 80 GB pipeline into 32 GB of VRAM. Context. MiniMax's H3 Max model topped video-generation quality charts in the month before this write-up, roughly September 2026. FastH3 was developed specifically to optimise H3 for cost, and the author notes that running H3-class models continuously can cost upwards of 6,000 USD per day. FastVideo's Preview V1 generated a 15-second, 1344 x 768 clip in 47.2 seconds on one NVIDIA B200, which the author describes as "much more expensive" than an RTX 5090. The gap between a data-centre accelerator and a 32 GB consumer card is where this workflow sits. How it works. The pipeline targets one RTX 5090 with 32 GB of VRAM and produces 448 x 256 clips at roughly 1x real-time. Six optimisations do the heavy lifting: four sampling steps per clip; VSA sparse attention that restricts full attention to 20 % of video blocks while still encoding the prompt; ClipProj, a NicoLab28 adapter that swaps H3's 32 B text encoder for a Qwen3-VL-4B, cutting encoder memory from about 15.7 GB to 4.5 GB; a pruned INT8 weight checkpoint from FastVideo; a fused FP4 MLP layer building on ByronLeeeee's work; and Kijai's INT8 video decoder that rebuilds frames in tiles, with NVENC encoding each finished tile while the decoder prepares the next. The workflow retains only the last decoded frame for the next clip rather than caching the full sequence. Combined, these reduce VRAM from roughly 80 GB to below 30 GB. Its read. The headline is "15 seconds in 15 seconds," but the resolution is 448 x 256. That is not a deliverable. It is a streaming placeholder at best. The 6,000 USD-per-day cost the author cites is real, but it describes continuous high-resolution H3-class inference on data-centre hardware. A rented RTX 5090 at roughly 70 cents per hour changes the unit economics by about two orders of magnitude, at 448 x 256. The SolRefinery upscaling pass (7x pixel density at 3x GPU cost) is the bridge to something presentable, yet the source calls it preliminary testing with no stated availability or licence. The more important second-order effect is architectural: tile-based decode with overlapping NVENC encode, plus single-frame carry-over, turns a batch job into a streaming pipeline. That pattern - decode tile, encode tile, discard - makes "live" generation on consumer hardware plausible, and it is not specific to H3. Any video model with a tileable decoder benefits. The question nobody is asking: at what clip length does retaining only the last decoded frame start to bleed temporal coherence, and is there a known upper bound on segment length? What this changes. If the studio already has an RTX 5090 or can rent one at 70 cents per hour, comfystreamerh3 is a drop-in ComfyUI node and the 15-second, 448 x 256 pipeline is testable this week without new infrastructure. The immediate use case is internal: prototyping prompts, testing model behaviour, building a live preview in a client dashboard. It is not a client-facing deliverable at that resolution. SolRefinery upscaling is the practical next step, but the source gives it as preliminary and names no licence, so it is not clear whether it is shippable. If the current GPU pool tops out at 24 GB, the workflow will not fit as described and there is no operational change on Monday. License. The source describes comfystreamerh3 as "released open source" but does not name a specific licence. The licences for FastVideo's FastH3 V2 checkpoint, ClipProj, the FP4 MLP work, and Kijai's INT8 decoder are likewise not stated in the brief. Check each repository's licence file before embedding any of these in a client-deliverable pipeline. Key takeaways. * H3-class video generation now runs at roughly 1x real-time on a single 32 GB consumer GPU, but only at 448 x 256. * Six stacked optimisations (4 steps, sparse attention, 4B encoder swap, INT8 weights, FP4 MLP, tiled decode) cut VRAM from roughly 80 GB to under 30 GB. * The 6,000 USD/day figure applies to continuous high-end inference; a 70-cent-per-hour RTX 5090 rental is a different economic bracket entirely. * SolRefinery upscaling (7x pixels at 3x GPUs) is the path to usable resolution but is unverified and unlicensed in the source. * No specific licence is named for the ComfyUI node or any of the optimisation components; "open source" is not a licence. Sources. video generation comfyui rtx 5090 gpu optimization ai hardware How this post was made. Drafted from clustered primary sources by the models below, then read, edited and approved by a human before it was published. The sources are listed in full at the end of the article. * Drafted - 2026-10-10 13:45 UTC * Independent sources - 1 * cluster pair - clef:27b * cluster label - gemma4:12b * radar brief - gemma4:12b * research brief - qwen3.8:27b * draft article - qwen3.8:27b * short script - qwen3.8:27b * seo pack - gemma4:12b * Run - editorial-20261010T133900Z

ComfyUI gets an AI agent. Reading Time: 2 mins read Comfy has released an AI assistant for its open source node-based platform for generative AI. Comfy Agent plans, builds and fixes ComfyUI workflows right on the canvas. You can describe what you want to make and the agent wires up the nodes, connects the models and debugs what breaks, while you stay focused on the visuals. It currently runs on the Comfy Cloud, a local version is coming soon. Key features: * Build while you edit, it works on the same canvas and sees the same assets you do * Answer anything you ask, and you can point it at any node, image, or workflow * Run skills: create your own or use public ones from the community * Curated knowledge for model recommendations and parameter tips Coming soon: * Local support, including creating custom nodes and model/node downloads * GPU detection and workflow recommendations * LLM model selection

LTX-2.5 lands open weights with native multi-shot - and it's actually fast. LTX (Lightricks' open-weights spinoff) released LTX-2.5 today with native multi-shot video generation, day-one ComfyUI integration, and a 6.8-second 720p clip on dual GB200. The fastest practical video model for indie builders just got meaningfully better. TL;DR: LTX released LTX-2.5 today - an open-weights video model that generates a 10-second 720p clip in 6.8 seconds on dual GB200s, ships native multi-shot generation, and integrates into ComfyUI on day one. It's the fastest practical video model available to indie builders, and the licensing terms (free under $10M ARR) make it immediately usable for most studios. LTX-2.5 is not the biggest name in AI video. It's not trying to be. The model comes from LTX, the open-world-model spinoff of Lightricks - a Jerusalem-based company better known for Facetune and Videoleap. Lightricks has been bootstrapped and profitable since the consumer-app days, which means LTX's incentive structure is different from the VC-fueled frontier labs. CEO Zeev Farbman is blunt about it: "We're definitely not doing this as philanthropy." The strategy is open weights with a revenue gate - free under $10M ARR, paid above. Today's release lands that strategy on the strongest technical ground yet. Key takeaways. * LTX-2.5 generates a 10-second 720p clip in 6.8 seconds on dual NVIDIA GB200 - fastest practical video model available * Native multi-shot generation solves character and scene consistency across cuts in a single pass * Day-one ComfyUI integration plus open weights under Hugging Face give indie builders immediate access What's new in LTX-2.5. The headline number is speed: LTX says a 10-second 720p image-to-video clip takes 6.8 seconds on dual NVIDIA GB200 hardware. That's faster than real time. The LTX API tier produces 1080p in 23.7 seconds - still faster than competitors on the same task. LTX's own benchmarks put Google's Gemini Omni Flash at 52 seconds for the same job, Veo 3.1 at 70 seconds for an 8-second clip, and MiniMax-H3 at 180 seconds. Treat these as vendor-reported; the same caveat applies to LTX's 67% win rate in blind preference tests, which the company itself labels preliminary. The technically more interesting feature is native multi-shot. A single generation produces multiple connected shots, with character, scene, lighting, visual style, and voice held consistent across cuts. This is the problem that has plagued every earlier T2V/I2V model - you could generate a single beautiful shot, but stitching shots together broke identity. LTX-2.5 generates the whole sequence in one pass. Other notable changes: * A new diffusion video decoder that reduces artifacts in high-motion footage and reconstructs fine detail like text and faces * A custom Gemma 4 language backbone with prompt enhancement for multi-subject scenes * A pretrained checkpoint tuned for physical AI and robotics - a base for teams to fine-tune on domain data that looks nothing like cinematic video * A substantially improved distilled model that delivers near-full quality at lower cost, optimized for NVIDIA RTX GPUs What this means for builders. The combination that matters: open weights + day-one ComfyUI integration + multi-shot + speed. Indie builders, game studios, and prototype teams can pull the weights today and run them locally on any GPU with at least 16GB of VRAM. The model supports text-to-video, image-to-video, and audio-to-video in both portrait and landscape. Two variants are available. ltx-2-5-fast tops out at 4K with 24/25/48/50 FPS support, optimized for speed and lower cost. ltx-2-5-pro caps at 1080p but is tuned for higher fidelity. Both support the automatic duration mode - send "duration": null and the model picks clip length from the prompt itself. The revenue gate is the part to watch. Most studios and indie teams fall under $10M ARR - they get free use. Larger companies negotiate a license. The gate means LTX can grow with its users without chasing VC returns, and it means builders can ship product without license anxiety in the early years. The competitive landscape is now structurally clearer. Closed API leaders (Veo, Kling, Gemini Omni Flash) are still ahead on raw benchmark scores, but open-weights alternatives are no longer a generation behind. For anyone who needs on-prem deployment, data control, or fine-tuning access, LTX-2.5 is now the default starting point. The take. LTX isn't chasing AGI or frontier-model rankings. They're shipping a video model that builders can actually deploy today, with the licensing structure to make it sustainable. The multi-shot feature alone is worth the upgrade for anyone doing narrative work. The speed numbers will hold up; the preference-test win rates are less certain. For most indie builders and studios, this is the model to start with. The closed-API alternatives only win when you specifically need their distribution, brand safety, or a feature LTX-2.5 hasn't shipped yet. Sources. Get the next briefing. Signal-first AI briefings, weekday mornings. One concise briefing with three signals, why they matter, and one action to take. Free. No spam. Unsubscribe anytime. · Weekday mornings.