@HclHno3 "enshittification" is the pop culture term for this.....but yea.... en.wikipedia.org/wiki/Enshittif…
English
DMBTrivia: Setlists
14.7K posts
DMBTrivia: Setlists
@DMBTrivia
Sorry we have discontinued this project due to lack of time. (Was: Realtime Setlist Updates -Free Android DMB Trivia & Setlist App
California 加入时间 Haziran 2012
214 关注6K 粉丝
AI video models have a shrinkflation problem.
Launch day: full model, stunning quality, one-shot results.
1 months later: silently swapped to a distilled model, worse outputs, you need 5 draws to get what used to take 1.
The price stays the same. The burger got smaller.
A 10-sec video costs $1.30–$5.00 in real GPU compute. Companies charge $0.05–$0.20. The math doesn't work — so they distill the model and pray you don't notice.
Sora did it. Seedance 2.0 is doing it. #Seedance2
English
@AiBattle_ It's not open source/weights. It's built on Waver which they never released.
ByteDance doesn't really release open weights.
English
@Lentils80 Where did they say this will be open weights? Unless they have specifically said that somewhere I don't think that's accurate.
It's built on Waver which the did not release open for local use.
English
@JustinLin610 Fingers crossed you still have music model on the way - appreciate all the open source work y'all do - truly outstanding.
English
@victormustar GLM is supposed to be releasing a new image model any day now: x.com/bdsqlsz/status…
青龍聖者@bdsqlsz
I guess it from GLM.
English
I've nearly ditched closed source models for image generation (in creative workflows).
For most tasks, the cost/speed ratio doesn't justify closed models anymore. Open models have caught up on quality and I can try 15 new images in 1-2 sec. Iteration speed matters.
Sharing the few models I currently use ⬇️
English
@FPCGameSoftware @Rainmaker1973 Easier would be get the model more training data......this is inevitable tbh
English
@Rainmaker1973 I wonder if people with stronger minds would be able to create clearer images?
English
@bdsqlsz @IamEmily2050 Doesn't seem terrible from a transformers update this morning. Claims it's out Jan 10.
English
@IamEmily2050 I think AR models will be heavy and difficult to run locally.
English
Interesting paper from KlingAI. Klear: Unified multi-task audio-video generation;
- a single 26B DiT (T2AV, TI2AV) w/ high-fidelity lip-sync;
- uses Omni-Full Attention & Random Modality Masking for training;
- comparable to Google's Veo 3.
arxiv.org/abs/2601.04151…
English
@CongWei1230 @bdsqlsz Need a ComfyUI node developer to take this one on. The top names in that realm are busy w LTX, so hoping someone who could implementing it (but may be reluctant assuming the usual devs pick it up) may give it a go. @ me if so please!
English
Thrilled to open-source UniVideo🎬!
UniVideo brings unified multimodal understanding, generation, and editing to the video domain
One framework for
• video/image understanding
• text/image → image/video generation
• free-form image/video editing
• reference-driven image/video generation/editing
Code: github.com/KlingTeam/UniV…
Model: huggingface.co/KlingTeam/UniV…
Project Page: congwei1230.github.io/UniVideo
Huggingface: huggingface.co/papers/2510.08…
English
@bdsqlsz Hope someone implements this into Comfy. I still love Hunyuan V1 and feel if its I2V model hadn't been such a bust it could still be relevant. This model addition seems to potentially add good I2V functionality. Might be too difficult to vibe code, but am going to try.
English
Kling teams drop new Open source Video model！
Cong Wei@CongWei1230
Thrilled to open-source UniVideo🎬! UniVideo brings unified multimodal understanding, generation, and editing to the video domain One framework for • video/image understanding • text/image → image/video generation • free-form image/video editing • reference-driven image/video generation/editing Code: github.com/KlingTeam/UniV… Model: huggingface.co/KlingTeam/UniV… Project Page: congwei1230.github.io/UniVideo Huggingface: huggingface.co/papers/2510.08…
English
@Sean60133791259 @bdsqlsz Right cause they're mutually exclusive? Personally not a fan of the LtX look, and I'm sure there are others who spent time with the OG versions (.9x) who immediately pick up on its flaws.
English
@ostrisai @kohya_tech Pretty sure @Kijaidesign's WanVideoWrapper has an option to do either or on the Lora loader.
English
I think all inference engines merge the LoRAs in. Merging into a low precision/quantized model tends to make the LoRA weights get rounded away. Keeping the LoRA on side chain would solve this, but adds additional compute. Using multiple LoRAs would increase compute even more. I would love a toggle option to keep the LoRAs unmerged with inference engines, but I don't think any have that.
English
画像・動画モデルの推論時のLoRA適用って、重みにマージするのが主流なんだろうか。学習時と同じように推論結果を加えれば、元モデルを量子化してもそれなりに効きそうだけど……。
日本語
@MiniMax_AI Would love training specific to ComfyUI in a future update. Claude is the only model that seems to have a really good grasp on coding nodes/dealing with the various model types. Appreciate your work!!
English
M2.1 works well as a cognitive core that orchestrates other programs/tools/agents.
It reasons efficiently and doesn't get distracted during long-running tasks.
And it's lightweight enough to run at home:
Awni Hannun@awnihannun
Running four simultaneous OpenCode agents works well with mlx_lm.server continuous batching and MiniMax M2.1 on an M3 Ultra:
English
English
@Alibaba_Qwen @ostrisai I can't make sense why you dont publish bf16 and fp8 scaled compatible with comfyui
I have been whole day trying to compile most accurate fp8 scaled version
English
Qwen-Image-2512 has been added to AI-Toolkit. Big thanks to @ostrisai
Ostris@ostrisai
Qwen-Image-2512 has been added to AI-Toolkit. Training a 3bit ARA for it now that should be ready in ~24 hours.
English
@sirflak @zoom_will @Ali_TongyiLab Was a hint an audio model by QWen was forthcoming by @JustinLin610 in late October, but sadly haven't seen any other mentions of this before or since. (Perplexityed/Groked even)
Can still cross our fingers for early 2026
English
@zoom_will @Ali_TongyiLab That would be nice. Right now we have Tencents Songgeneration which they are cooking v1.5 rn. x.com/cocktailpeanut…
cocktail peanut@cocktailpeanut
100% Free Suno Alternative on Your PC with 1 Click. SongGeneration Studio, by @Bazed_X_, is a web UI that generates up to 4m30s high-quality songs with just 10GB VRAM, powered by Tencent AI Lab's SongGeneration model. And you can run it on your PC with 1 click.
English
our new model is actually not bad even in comparison with the sotas. coming soon!
Chen Cheng@cherry_cc12
@craftian_keskin Qwen
English
@oyacaro @wildmindai These are from the 20b - takes forever even on top HW w/ the distilled models, but people saying the quality isn't good haven't used it.
English
Surprise. Kandinsky 5.0 Pro beats Wan2.2 on the T2V bench: 19B model is on par with Luma Ray 3;
the 2B Lite outperforms the first version of Sora.
English
@virtualfilmer @wildmindai What? It's extremely heavy to run, but if you think the quality is bad you're not running the model properly. Most discussion is on discord in Banodoco.
Ex:
English
@wildmindai This can’t be right. Kandinsky’s results are… not SOTA. They are, in fact, quite poor.
English
@multimodalart @huggingface Is this gone for good now? Says paused. :(
English
PRO for PROs
Nano Banana PRO is available at no cost for @huggingface PRO subscribers on Spaces, go bananas 🍌
English