AI Radar

12 madde 28/31 kaynak 198 tarandı
düşen kaynak: import-ai, reddit-localllama, reddit-aivideo

Üretim4

Did a quick anime-to-realism test in MiniMax H3 with just a video reference + lora

görselvideo·reddit-stablediffusion·3 saat önce

The post at the bottom inspired me to try a simple anime-to-realism test with MiniMax H3, so I figured I’d share a few results. I didn’t generate any keyframes or stills. I just fed the original video directly into MiniMax H3 as the video reference, applied the lora, and basically told it to convert it to live action. The output quality was honestly better than expected. Not saying it’s perfect or amazing or anything, but for such a simple setup (ComfyUI official Minimax H3 Ref2V workflow + lora) I was pretty impressed. I didn’t use any character reference sheets, so character consistency…

7.69
corroboration+0.00
kaynak+1.50
eksen+4.34
sinyal+0.00
tazelik+1.25
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.59

Portable .char consistent characters(same face, body, cloths & voice) - MinimaxH3, Flux2 & Krea2

görselvideo·reddit-stablediffusion·5 saat önce

Hey guys, I have been working on the consistent character portable format (.char) for a while & now there's a community characters page where you can grab a ready-made character and use it straight in ComfyUI. Same face, cloths & body across every video generation, and with a few seconds of voice sample the character talks in a consistent cloned voice with lip sync (MiniMax H3). Community characters Aron: https://www.omnichar.org/characters/aron You can use these characters via comfyui, omnichar app & cloud. Detailed guide on building full consistent character(workflow included): Previous…

7.62
corroboration+0.00
kaynak+1.50
eksen+4.34
sinyal+0.00
tazelik+1.18
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.59

VIS-Ground: Video Interactive Storytelling with Contextual Grounding

video·hf-papers·7 gün önce·63 oy

Video interactive storytelling enables viewers to actively steer how a video unfolds. However, once we allow viewers to intervene during generation, a new challenge arises: The viewer's request can have latent dependencies on both the grounding source and the current rendered video state. These dependencies may not be explicitly stated in any individual input, but emerge only when the source, rendered history, and new viewer intent are considered jointly. Existing interactive video generation systems primarily emphasize following viewer instructions, while source-grounded video generation…

6.25
corroboration+0.00
kaynak+2.00
eksen+3.30
sinyal+0.74
tazelik+0.06
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.16

Scaling Agentic Data for Long-Horizon Terminal Intelligence

görselmodel3D·hf-papers·5 gün önce·129 oy

Terminal agents are moving beyond local operations toward completing real-world professional tasks. Short trajectories can demonstrate commands and tool calls, but they do not fully show how agents accumulate evidence through exploration, use it to inform judgments, continually adjust strategies based on feedback, and turn intermediate results into final deliverables. High-quality long-horizon trajectories capture these agentic behavioral patterns and their connections within a single problem-solving process, providing richer experience for sustained autonomous work. However, environment…

5.15
corroboration+0.00
kaynak+2.00
eksen+1.76
sinyal+0.86
tazelik+0.06
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.47

Oyun4

Video game industry believes EU Kids Act will drive small creators and old games into extinction

oyunvideo·reddit-gamedev·5 saat önce

A new law aimed at protecting children could kill off old video games and threaten small developers, industry trade groups have warned. The EU Kids Act was unveiled by commission president Ursula von der Leyen last month as a chance to “give childhood back to our children” and put parents “in the driving seat”. It would require videogames to be catered to children by default with stringent requirements including an eight-hour nighttime lock-out, a ban on screenshots, and guardian supervision. Gamers wanting to play the full version of games would have to pass age verification checks to access…

6.54
corroboration+0.00
kaynak+1.00
eksen+4.34
sinyal+0.00
tazelik+1.20
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.00

Defender’s Advantage is Bad for Player Agency Strategy Games.

oyun·reddit-gamedev·az önce

Strategy games are all about the decisions you can make as a player. Players are also generally very good at identifying the short-term costs and benefits of a decision. So when you want to encourage chaotic board states in a strategy game, the last thing you want to do is discourage players from attacking. If you can throw aside the simulation lens for a moment, what I hope to convey is that as strategy game designers, we should evaluate attacker/defender advantage as an entire system within our game that can strongly affect player behavior. Player agency is probably the most important…

5.89
corroboration+0.00
kaynak+1.00
eksen+3.52
sinyal+0.00
tazelik+1.37
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.00

Epic Games delays Disney game and tells developers to go "AI first"

oyun·gamesindustry·19 saat önce

Epic Games has delayed its upcoming Disney collaboration - a shooter featuring characters from the Marvel and Star Wars franchises - and told developers to move to an "AI first" production model. Read more

5.27
corroboration+0.00
kaynak+1.00
eksen+3.52
sinyal+0.00
tazelik+0.75
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.00

Modeller & Araştırma3

Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

↑ yükselişte · yenidenHN'de GPT-6.1 0→8 hikâyemodel·openai·4 gün önce

Using GPT-6 Astra in Codex, Asana made its browser agent 76x cheaper and 5x faster in tests to offer customers more capable models.

7.91
corroboration+0.00
kaynak+3.00
eksen+2.86
sinyal+0.00
tazelik+0.06
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+2.00

AI is a tool and using it well takes skill

↑ yükselişte30 oy/saataraçmodelşirket·hackernews·17 saat önce·503 oy
7.33
corroboration+0.00
kaynak+1.50
eksen+1.43
sinyal+1.10
tazelik+0.80
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+2.50

Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB

model·marktechpost·6 saat önce

Z.ai’s GLM-Legal is a legal AI model built on GLM-5.3-Flash. It scores 19.2% all-pass on Harvey’s Legal Agent Benchmark, nearly 3x its base, at $0.15 per test. Access is by application. The post Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB appeared first on MarkTechPost .

4.84
corroboration+0.00
kaynak+0.50
eksen+2.86
sinyal+0.00
tazelik+1.14
capraz_kaynak+0.00
gurultu_cezasi+0.00
hype+0.34

Ekosistem1

Kısa kısa8