Jasper de Bruin

@huggingface All of my (private) organisation HF spaces now fail to 'start'. After some idle time during the weekend, none of them appear to boot anymore. No clear error, nothing changed in the meanwhile. Any clue on how to debug or resolve this?

@OfficialLoganK any idea when gemini 3.1 pro and flash will become generally available? We can't use it in production with current rate limits in Vertex

@huggingface account registration is showing a 500 error

@BrengOV @Arriva_NL Zie deze conversatie

@JdeB90 @JdeB90 Goedenavond, goed dat je dit navraagt! In het weekend wordt deze treinverbinding niet door ons verzorgd, maar door Arriva. Je vindt dit terug op o.a. 9292.nl (obi41.nl/2p8j2cwn). Hopelijk kunnen onze collega's van Arriva je verder informeren. ^VM

@BrengOV Trein naar Arnhem rijdt net met een vaart Doetinchem De Huet voorbij. Wat is dit nou weer voor iets waardeloos? In de middle of nowhere, en de enige optie is weer een half uur wachten...

@joshwoodward @GeminiApp I way to organize my gems in folders would be nice. Also adding more customizability to the icon of a gem. Having multiple gems with the same letter and color icon isn't too great.

After Google IO, we ran our first-ever "Papercut Week" on @GeminiApp and fixed 84 issues. This morning, the winners were announced and will receive the coveted "Golden Bandaid" awards! Improvements you'll see: ✅ Better mobile swipe gestures ✅ Dynamic input fields for longer text ✅ Streamlined file upload flows ✅ New motion animations …and more Are you hitting other annoying papercuts in the app? Let me know, and we'll queue them up for fixing.

@eliasroman @avdnoord @OfficialLoganK Are there plans for LoRA (and possibly controlnet) support for Imagen and/or Veo?

@OfficialLoganK Will it be possible to finetune or create LoRAs for imagen 3 via GCP?

For now, you can use ImageFX, hang tight for AIS: labs.google/fx/tools/image…

Imagen 3 is now ranking #1 on the lmarena text-to-image leaderboard, and will be available in the Gemini API and AI Studio soon : )

@Lykon4072 @dango233max @kasukanra My LoRAs (mostly lokr) for SD3.5 (both M & L) have a negative impact on anatomy and concept bleeding. What is wise to reduce said issues etc? I can think of: - lower LR more steps - lower rank / alpha more steps - adding regularization images

@OfficialLoganK LoRAs for imagen and veo would be amazing. Upload a dataset (optional autotagging with gemini), train in gcp and host it with model garden

Native image output with Gemini 2.0 Flash is what I’ve always wanted from an image model, the consistency when iterating is so magical. All without any complex orchestration or develop tooling.

@Lykon4072 @dango233max @kasukanra This is new to me, can you point me to some info somewhere?

@JdeB90 @dango233max @kasukanra you need a restoration vector, like a "bad" lora applied in negative weight.

@Lykon4072 @dango233max @kasukanra What are your current best practices for SD3. 5 finetuning/LoRA training? LyCORIS (lokr) with SimpleTuner seems to work best for me but anatomy gets worse with lokr even when using SLG

@dango233max Noticed SLG is now added in Diffusers, but still patched weights are welcome. Do you know anything about an ETA?

@JdeB90 Some weight patchs are on their way Meanwhile-SLG is your friend ;)

To all fellow #SD35 tuners: If you have experienced difficulties in baking details into the model, please try the patched Kohya sd-scripts sd3 branch: * The default timesteps scheduling (Uniform) should be more SFT friendly and attend to details * More adjustable parameters now

@Lykon4072 @kasukanra I have been experimenting with SD3.5 style LoRA training based on the settings from the @StabilityAI Notion guide. The model learns the style but the anatomy seems to break (same prompt/seed without LoRA the anatomy is fine). Any ideas?

@Lykon4072 Thank you and @StabilityAI The results look amazing, combined with different sized models and the permissive license, you guys have outdone yourselves 🙏

@JdeB90 Hold your seat

@Lykon4072 any news on the upcoming release? I've got so many questions. ETA, Will it be 2B, are you planning on releasing bigger weights, how are results currently compared to SD3M? 🧸

@Lykon4072 Any news on the SD3.1 release yet? I can't wait to try out the new weights!

@Lykon4072 @StabilityAI Awesome, thanks for the update! Any estimation of the release date? Even something vague like within 2 months or next year?

@JdeB90 @StabilityAI keep in mind 3.1 is still in development. If at some point we decide to do something crazy like double the size, nothing will be compatible😄 So before people accuse me of lying, I need to specify "at the current stage"

@StabilityAI @Lykon4072 Do you have a ETA for the updated SD3 weights? Also any idea if controlnets / LoRA support will be added?