This time it's more about tools than new models. On October 6 ComfyUI 0.39.0 came out: among the new features are 16-bit float EXRs, handy for compositing, and an option to work truly offline. If you want to train your own LoRAs, there's Fizgig v7, which now also does full fine-tuning starting from 8 GB, and LoRAlab, which puts nine trainers behind a single button. We wrap up with VNCCS 3.2, for creating consistent characters with Qwen-Image 2.1.
In brief
- ComfyUI 0.39.0: 16-bit float EXRs, videos saved at higher quality, the new
--offlineand--disable-partner-nodesflags and lighter VAEs for MiniMax H3. - Fizgig v7: SDXL and Anima arrive, plus full fine-tuning for six models; Qwen-Image 2.1 trains from 12 GB.
- AcademiaSD LoRAlab: nine LoRA trainers in a single install, from 4–8 GB of VRAM, free and open source.
- VNCCS 3.2: consistent character sprites in ComfyUI with Qwen-Image 2.1 and MiniMax H3, even with a transparent background.
ComfyUI 0.39.0: 16-bit EXR, offline mode and cleaner videos
A week after 0.38, late in the evening of October 5, the stable 0.39.0 release of ComfyUI came out. We had already seen many of the new features pass through the development branch, like the FLUX 3 Image node, Eleven v4 and the fixes to resource scanning. Now, though, they're in an official release, the one that also reaches people updating the portable version. Here's what we think matters most for creators:
- 16-bit float EXR. The Save EXR node now saves in half float by default. If you take your images into After Effects, Nuke or Fusion, that means much lighter files than 32-bit, with enough range for compositing.
- Better saved videos. Save Video now uses higher quality settings: CRF 18 in H.264 and 24 in AV1. Files will be a bit heavier, but with fewer compression artifacts.
- Truly offline. The old
--disable-api-nodesflag did two things at once: it hid the paid nodes and blocked every request from the interface to the internet. That's why the new Nodes Manager stayed empty. Now there are two separate flags:--disable-partner-nodesonly hides the API nodes,--offlineshuts everything off. The old flag still works, but it's deprecated. - "Light" VAEs for MiniMax H3. ComfyUI now also loads slimmed-down VAEs like LynnReal light VAE, with a ready-made conversion by Kijai. The license is the same as H3, so the European Union exclusion still applies.
- Fixes. The Qwen-Image 2.1 int8 and int4 cache no longer crashes, some memory issues with the MiniMax VAE are fixed, and pinned memory switches itself off on integrated GPUs.
On the paid nodes side, Ideogram 4.5, HeyGen Video 1.0 and Grok Imagine Video 1.5 Lite arrive. If you use Luma Ray 2 through ComfyUI, note that those nodes are deprecated and the endpoints will be removed on October 24. Our verdict: it's an update worth doing. As always, though, make a copy of the folder or the Python environment first, especially if you use lots of custom nodes.
Fizgig v7: full fine-tuning even on 8–12 GB cards
We covered Fizgig at the end of September, when it brought Qwen-Image 2.1 LoRA training to 12 GB cards. On October 5 version 7 came out, and it's a big jump. On top of LoRAs, it now does full fine-tuning of all six supported models: FLUX.2 Klein 9B, Krea 2, Qwen-Image 2.1, MiniMax H3 and the two newcomers, SDXL and Anima. The result is a complete model you load in ComfyUI in place of the base one.
The trick to fitting it in little memory is training the model one piece at a time and keeping the rest in system RAM. According to the author's table you need:
- 8 GB for SDXL and Anima (with 10 GB the whole model trains at once, faster);
- 12 GB for Qwen-Image 2.1;
- 16 GB for Klein 9B, Krea 2 and MiniMax H3.
So on a 12 GB RTX 3060 like ours, on paper, fine-tuning SDXL, Anima and Qwen-Image 2.1 becomes possible. With SDXL you can start from any checkpoint: by default Fizgig downloads Juggernaut XL v9. Two caveats. SDXL and Anima are marked as experimental, and Anima has a non-commercial license. Fizgig itself is open source under the Apache 2.0 license. To see how it works there's the video "What is Fizgig - LoRA & Fine-tune Studio?" from the ShootTheSound channel, run by the author himself.
LoRAlab: nine LoRA trainers in a single install
If you've tried training LoRAs for different models you know the drill: every trainer has its own Python environment, its own dependencies and its own interface. AcademiaSD LoRAlab Trainer Studio brings it all together. With a single install and a single launcher you get trainers for Qwen-Image 2.1, Krea 2, Z-Image, FLUX.2 Klein 9B, Ideogram 4, Anima, SDXL, LTX-2.3/2.5 and MiniMax H3. ComfyUI Wiki has covered it too.
Models are loaded in 4-bit (NF4), and the text encoder and VAE run only once, before training. That keeps the graphics card free for training. According to the README, 8 GB is enough for Qwen-Image 2.1, Krea 2, Z-Image and even MiniMax H3, while SDXL and Anima in NF4 need just 4. Klein, Ideogram 4 and LTX need 12 GB. There's also an automatic captioner, previews during training and one-click sending of the LoRA to ComfyUI. For models that support it you can also train editing LoRAs, the ones that learn a "before and after".
The project is new (it started on September 28), it's free and MIT licensed. On October 6 it added the rsLoRA, LoRA+ and LoKr modes. The author himself says there may be bugs and default settings that need polishing: for us it's a great candidate to try, but on test projects, not on a job with a deadline. To see it in action there's the Spanish-language video "Entrena LoRAs para Qwen-Image 2.1, incluso LoRAs de edición" from the Academia SD channel. Remember that LoRAs inherit the license of the base model: Qwen-Image 2.1 has a "research" license, worth checking if you work for a client.
VNCCS 3.2: consistent characters with Qwen-Image 2.1 and MiniMax H3
VNCCS is a ComfyUI node suite that creates "sprite" sets: the same character in different poses, outfits and expressions. It's built for visual novels, but it works for mascots, storyboards and indie games too. Version 3.2, released on October 3 and already updated to 3.2.3, switches engines. The default model is now Qwen-Image 2.1, in its INT8 version, with the fast 6-step Viggle Turbo mode built in. Alternatively you can use MiniMax H3: VNCCS generates a short video and keeps the first frame as the character image.
Other new features include sprites with a transparent background, a rebuilt Character Creator with a choice between half-body and full-body framing, and a much larger style library, which according to ComfyUI Wiki reaches 358. The nodes are free (MIT license), but the model licenses apply: Qwen-Image 2.1 is "research" and MiniMax H3 excludes the EU. So if you're in Italy, the Qwen route remains the most sensible one.
Try it now
If you work with After Effects, after updating to ComfyUI 0.39.0 try replacing the final PNG in your workflow with the Save EXR node. It now saves in 16-bit float by default. Import the file into After Effects in a 16- or 32-bit project and try a heavy color correction, for example +2 exposure or an aggressive curve. Then compare the result with the same PNG: in the shadows and skies you'll see the difference right away. And if you launch ComfyUI with the old run_nvidia_gpu_disable_api_nodes.bat and the Manager looked empty, now you know why. Just use --disable-partner-nodes instead of --offline.
To choose which model to train your first LoRA on, read our guide to AI image models in ComfyUI. For video there's the one on AI video models in ComfyUI. If you're thinking about a card with more VRAM for training, take a look at our guide to the best GPUs for AI. If you're starting from scratch, begin with the ComfyUI beginner's guide.
That's all for now. Download our free workflows from the ComfyUI section and follow us on the YouTube channel @tutorial3d. See you at the next update with more AI news.
