tarun bana
554 posts
マルチショットでの背景の一貫性・整合性は皆さんどうしています？polyhavenの360度lon-lat画像を入れて検証しました。ファーストフレームだけはNano-BananaProで作り、キャラはElementsとキャラクターシート、lon-lat画像はimageリファレンスにして、同じプロンプトでSeedance2.0とKling3.0 Omniで生成しました。Klingでも360度lon-lat画像と認識させるプロンプトがあるのか？ちなみに360度lon-lat画像だけだとSeedanceもKlingも常に背景が歪んだので歪みのないファーストフレームを入れました。
#Seedance2 #klingai #aivfx
日本語
@peterom NBP is soo good, it's not possible to use anything else, once you use it. It's that simple. Trust me on this one. Most people don't stretch these models, NBP is like science fiction.
English
Flux.2.Klein 4b/9b are very good models that haven't gotten the reception they deserve due to residual bad vibes from Flux.1.dev's distillation/aesthetic, the poor initial Flux.2 landing, + overshadowing from NBP.
But they're v. capable + responsive to training!
English
@Yokohara_h My main question is control, did u want the video to come out like this? I feel like this model has good output, but it does what it wants.
English
ComfyUIでseedance2.0使えるようになったから試した
フォルダに残ってた適当な画像を動かした
しかしseedance2.0触るたびに今までの映像制作について考えさせられるなあ
まあ値段が高すぎるけど・・
日本語
@cloneofsimo Not gta 6 but I'm trying to make like superpowers, where rules interact. And it just stops being usefull as the task gets complicated, tho I am getting very far. Note- I don't use any harness currently. I use only chat and VScose.
English
In this chart below, I've evaluated quantized versions of Gemma 4 31B + the original bf16 model (all thinking disabled):
Intel/gemma-4-31B-it-int4-AutoRound
RedHatAI/gemma-4-31B-it-FP8-Dynamic
google/gemma-4-31B-it
nvidia/Gemma-31B-IT-NVFP4
RedHatAI/gemma-4-31B-it-NVFP4
cyankiwi/gemma-4-31B-it-AWQ-4bit
Can you guess which bar is the original one (bf16)?
English
@MaziyarPanahi Insane levels of surveillance tech, i guess u still need the satellite, but government already got that
English
Gemma 4 analyzes the video. Generates key questions. Calls Falcon Perception.
"Find all the people." 156 found.
"Detect only white cars." 8 found.
A 26B model is running agentic multi-QA vision orchestration.
The models are running locally on a MacBook with MLX. No API.
English
@bnjmn_marie Can't wait to see this. I feel like gemma 4 should be much better, in my test it's the only model that has passed it that's not paid and from big 3 main labs. I hope there's a section where you have your own opinion too, beyond the benchmarks. Based on your testing.
English
Gemma 4 31B vs Qwen3.5 27B, Thinking Enabled
I ran multiple benchmarks multiple times. Gemma 4 31B looks better and more stable (smaller accuracy variations between runs, which makes sense since it generates shorter sequences).
I'll publish my full results and analysis on my blog later this week (link in profile).
English
Kogar and the Old Man. A few weeks ago, Facebook’s algorithm showed me an epic illustration by John McKenzie. His style is quite unique, so I recommend checking out John’s account to appreciate it.
English
@Kazi5isAlive Yeah, but image is also generated using text, or are u making images in photoshop ?
English
@0xCVYH I read through the hugginfacve page, i don't get it. I could ask llm, but I'm sure those dumb things will just hallucinate a explanation. What is this? I'm guessing something to do with low vram or faster with same vram something like this.
English
quantizei o wan2.2 animate 14B pra video-to-video com hlwq. modelo de animacao de imagem, 8-bit, apache 2.0.
quase mil downloads no hf sem eu ter divulgado antes. base e o wan-ai oficial.
huggingface.co/caiovicentino1…
Português
@Sauers_ Yup, same. And my use case is not even that complex. But same experience, it needs harness, well I'm guessing. I'm gon know soon I'm building harness now.
English
A few weeks ago I had GPT 5.4 derive and prove dozens of genetics theorems over multiple days. It did manage to write correctly-compiling math, but there were subtle improper assumptions or specifications leading to the results being useless and disconnected from reality
Mira@_Mira___Mira_
2 days ago GPT 5.4 ran for 12 hours proving 107 different theorems while I was sleeping. Yesterday it ran for 14 hours formalizing Shor's Algorithm. I think that needs another day or two to finish.
English
@iabom When running locally, most people use quants, so when testing a model that is mostly run locally, it makes sense to use a quant. That is my reasoning anyway.
English
Gemma 4 31b scores 52.3% and is the strongest open model on on WeirdML, ahead of GLM 5 and gpt-oss-120b. This score is comparable to o3 and gemini 2.5 pro, and well ahead of qwen 3.5 27b at 39.5%.
Gemma 4 is also significantly cheaper than other models with the same score.
I ran this locally through ollama, with a 4-bit quant (q4_K_M). Full precision might score even better. The costs assume $0.14/$0.40/M.
Håvard Ihle@htihle
WeirdML v2 is now out! The update includes a bunch of new tasks (now 19 tasks total, up from 6), and results from all the latest models. We now also track api costs and other metadata which give more insight into the different models. The new results are shown in these two figures. The first one shows an overview of the overall results as well as the results on individual tasks, in addition to various metadata. The second figure shows cost vs performance and shows a clear scaling with better results for higher costs. We also have a very varied pareto frontier with 11 models from 6 different companies having the best accuracy for a given cost for at least some of the cost range. Grok 3, Claude Opus 4 and GPT 4.5 are the ones that underperform for their costs, while Gemini pro and o3 pro have the best results at the highest costs. Qwen3 30B3A, grok 3 mini and deepseek R1 also each represent a good chunk of the pareto frontier.
English
@francoisfleuret Anything verifiable and instant output that can be evaluated= AI good. Anything else = horseshit covered in gold foil. People actually take advice from these things. LMAO
English
> LLM write a very convincing explanation hinging on [WHATEVER]
> me asking why [WHATEVER] would make any fucking sense
> LLM apologizes and goes into self-criticism session
English
Most U.S. generative media AI labs that raised hundreds of millions or billions seem to have quietly given up on building the best foundation model---They pivoted to consumer apps---Is there any AI lab left, actually focused on building the best foundational generative AI model?
English
@Yokohara_h yes, and not only that, it all looks so same. There is a sort of AI likeness, not that its low quality. just not distinct enough in the art somehow. Its the same thing that happens to every Unreal engine 5 game, they all just feel same somehow.
English
seedance2.0とか動画生成どれも値段高い問題
これじゃあ小金持ちのおっさんくらいしか動画生成で遊べないよね
日本語
there is a very good chance you will be beating it to a generated woman in a few years. bcz real stuff's dopamine will be lower than the generated perfection.
English