The most useful news if you generate video locally comes from NVIDIA: LongLive-Plug, a set of LoRAs that brings Wan and MiniMax H3 down to just a few sampling steps. On the 3D side, Blender 5.3 shows off its native Gaussian splats and the DLSS denoiser in the Cycles viewport, while the beta slips by a few days. And ComfyUI gets 16-bit EXR, very handy if you do compositing.
In brief
- LongLive-Plug: NVIDIA LoRAs to generate video with Wan 2.1, Wan 2.2 5B and MiniMax H3 in about 4 steps instead of 50.
- Blender 5.3: Gaussian splats in Workbench, EEVEE and Cycles, DLSS in the viewport; the beta is expected early next week.
- ComfyUI: 16-bit float EXR saving, partner nodes for Ideogram 4.5 and HeyGen Video 1.0.
- ComfyUI-RMBG 3.2.0: cutouts with SAM 3.1 and a switch to free VRAM right away.
- Consistency LoRA for Qwen-Image 2.1: edits that don't shift the framing.
- Comfy Dev Platform Challenge: from October 5 to 19, with $10,000 and three RTX 5090s up for grabs.
LongLive-Plug: Wan and MiniMax H3 in a few steps
NVIDIA has published the LongLive-Plug collection on Hugging Face: six LoRAs, two for each base model. One reduces the sampling steps (few-step), the other "bakes" CFG into a single pass. The supported models are Wan 2.1 T2V 14B, Wan 2.2 TI2V 5B and MiniMax H3. According to the project page, using both LoRAs together takes you from 50 to 4 steps.
Why it matters to you. On a 12 GB card like our RTX 3060, the problem with video models isn't just memory: it's time. Every test clip costs minutes, and ten tests cost a morning. Cutting the steps tenfold changes the way you work: you can explore shots and movements before launching the final version. The interesting idea is that the LoRA strength becomes the guidance control, so you adjust it through the adapter weight.
Our take. For our cards the natural candidate is Wan 2.2 TI2V 5B, the lightest of the three. The Wan LoRAs are under the Apache 2.0 license, so you can use them for work too (the base model's license still applies). The MiniMax H3 ones follow MiniMax's community license instead, the model card advises against using the two LoRAs together for now and, according to ComfyUI Wiki, they need to be converted before you use them in ComfyUI. It's a "research demo": expect some quality loss compared to 50 steps.
Blender 5.3: Gaussian splats and DLSS, but the beta is running late
In episode 290 of Blender Today LIVE (the official Blender channel) Pablo Vazquez rounded up the new features merged into the main branch. The most anticipated one: according to the release notes, in Blender 5.3 Workbench, EEVEE and Cycles can render Gaussian splats, imported from PLY, SPZ or USD as regular point clouds, so also inside Geometry Nodes. The same notes admit that performance is "not ideal" and, as Digital Production sums up, export and full color management are still missing.
If you have an NVIDIA card there's also a new DLSS denoiser for the Cycles viewport: a cleaner and faster preview, because it reuses samples from previous frames. Be careful though: it's active in the daily builds but requires an NVIDIA driver that hasn't been released yet. As for the beta, it was scheduled for September 30, but it hasn't been announced yet: in the Render & Cycles module meeting they talk about "early next week".
Our take. Splats are the natural bridge between AI and 3D: you capture (or generate) a scene and use it as a real background, with a free camera. Having them inside Blender without add-ons is a nice step, but for client work we'll wait at least for the beta. If you want to see them in action, "Blender 5.3 Got Gaussian Splats!" by CGMatter and "Gaussian Splats in Blender 5.3!" by Cartesian Caramel are two good starting points.
ComfyUI: 16-bit EXR and two new paid services
On October 1 PR #16706 was merged into the ComfyUI code (master branch, so not yet in a stable release): the SaveImageAdvanced node can save 16-bit half float EXR files. 32-bit remains the default. If you bring your images into After Effects, Nuke or Fusion, this is good news: 16-bit float is the de facto standard in compositing and takes up about half the space. It ties in nicely with the color features (ACEScct and LogC3) we covered in the past few days.
Partner nodes have also arrived for Ideogram 4.5 (generation, editing and precise editing) and for HeyGen Video 1.0 (video from image, from text or from references). These are pay-per-use online services: prices aren't listed in the PRs. Finally, a fix for MiniMax VAE offloading, useful precisely if you're short on VRAM.
ComfyUI-RMBG 3.2.0: easier cutouts with little VRAM
The ComfyUI-RMBG node pack, one of the most popular for removing backgrounds and segmenting objects, has reached version 3.2.0. What's new: a SAM 3.1 Multiplex node that loads safetensors files and selects subjects by name or with points, an "unload_model" switch on all the main nodes to free VRAM as soon as they're done, a resolution selector, an image comparison node and a much faster startup.
Our take. The switch to unload the model looks like a detail, but on 12 GB it makes a difference when you put a cutout and then a heavy model like Qwen-Image or Wan in the same workflow. The nodes are under the GPL-3.0 license; the models they download (RMBG-2.0, SAM, BiRefNet…) each have their own license: check it before using them for a client.
Qwen-Image 2.1: a LoRA that keeps the framing still
If you use Qwen-Image 2.1 editing, you know the problem: you ask "turn it into a watercolor" and the subject shifts by a few pixels, so you can no longer overlay the result on the original. ausboss's Consistency LoRA is made exactly for this: no trigger word, strength 1.0, two versions (step 1500, recommended, and step 2000, more precise but with slightly duller colors). The author tested it at around 1 megapixel and states that it doesn't strengthen weak edits: it only keeps the composition still.
Our take. For motion graphics it's extremely useful: an image that stays "in register" can be masked, animated or blended with the original. Watch the license: it's the Qwen Research License, the same one listed for the base model, which you should read carefully before any commercial use.
Comfy Dev Platform Challenge: three RTX 5090s up for grabs
On October 1 Comfy announced a developer challenge running from October 5 to 19: recreate in open source features that you pay for today, such as cinematic camera moves, character consistency or relighting. Up for grabs are $10,000, three RTX 5090s and API credits; the jury is made up of developers from ComfyUI, NVIDIA and Runpod. If you build workflows, it's a chance to get noticed (and the first 100 entrants get free credits).
Try it now
If you already have Wan 2.2 TI2V 5B in ComfyUI, download the two LongLive-Plug LoRAs for the 5B (few-step and the matching CFG one from the collection). Load them one after the other with the ratio recommended on the model card, 1 for the few-step and 0.5 for the CFG, set 4 steps (since the guidance is already in the LoRA, start from CFG 1) and compare the clip with the one from your usual workflow using the same seed. Keep an eye on the time per clip: that's the number that matters on a 12 GB card. Do this test on a copy of your workflow, because we haven't checked yet how the standard LoRA loader reads these files.
To pick the right video model, read our guide to AI video models in ComfyUI and, if you're weighing whether working locally makes sense, our comparison AI video: better locally or online?. If you're starting from scratch, begin with the beginner's guide to ComfyUI; for the graphics card there's our article on GPUs for AI.
That's all for now. Download our free workflows from the ComfyUI section and follow us on our YouTube channel @tutorial3d: see you at the next update with more AI news.
