TRGNRD

[TRGNRD @TrigonRed]

@ostrisai Haha 😄 If you want to be the best, be the best at everything. Trump ©

What do you think about training a LoRA/LoKR on prompts that trigger the safety filter, using matching images in different aspect ratios, then applying that LoRA with negative weights?

[TRGNRD @TrigonRed]

@hsdhcdev @SlipperyGem I've run plenty of tests. The model is really strong, but the safety block needs to go
The issue almost certainly isn't in the text encoder it's embedded in the DiT itself. Some manipulations already reduce its aggressiveness, and even a small finetune could be enough to break it

[TRGNRD @TrigonRed]

@SlipperyGem @hsdhcdev Actually, there are several approaches worth trying:

1. Fine-tune the model or train an NSFW LoRA/LoKr to shift its priorities.
2. Train a LoRA on censorship artifacts and apply it with a negative weight.
3. Identify where the censorship is encoded and selectively remove it.

[TRGNRD @TrigonRed]

@SlipperyGem I believe Krea 2 will be great and won't come with this kind of nonsense.

[TRGNRD @TrigonRed]

@SlipperyGem True.

I've tested Ideogram quite a bit, and I genuinely don't understand why they crippled the model like this. The funniest part is that it'll block completely harmless prompts, yet a long, detailed prompt can sometimes generate nudity just fine. The whole thing feels absurd.

[TRGNRD @TrigonRed]

@JulienBlanchon

[TRGNRD @TrigonRed]

@reve Ideogram 4 went open source. Krea 2 is just around the corner. Care to join the party? 😏

[TRGNRD @TrigonRed]

@bdsqlsz The model looks promising, but the safety system is absurdly aggressive. The safety filter can be triggered by something as simple as "an apple."

It's hard to understand why anyone would cripple such a promising model. The community will break through those restrictions anyway.

[TRGNRD @TrigonRed]

@ComfyUI @ideogram_ai Is this even real?

TripoSplats, now Ideogram 4, Krea 2 around the corner...

Maybe I slipped in the shower and this is all just a dream???

[TRGNRD @TrigonRed]

@angrypenguinPNG No words can describe the anticipation.
Go nuclear!🔥

[TRGNRD @TrigonRed]

@davidpwalter Time matters
In my opinion, the core idea here isn’t just viable. It’s probably the only genuinely reasonable and forward-looking direction for engine development at the moment.

[TRGNRD @TrigonRed]

@davidpwalter This is phenomenal work! We really need native support for Nanite-like technology in Godot. In my opinion, this is one of the most important things the engine should be focusing on today.

[TRGNRD @TrigonRed]

@3DxDEV7 @dante_leoncini I installed Blender on an HTC P3600 PDA back in 2007 or 2008 )))

[Stable Diffusion Tutorials @SD_Tutorial]

Nvidia's PID (Pixel diffusion Decoder)

It decodes latent representations into high-resolution images, replacing the decode–then–super-resolve cascade while achieving lower latency and higher visual quality

page:
https://t.co/kBhlc63uV2

HF repo:
https://t.co/IBgBJvG7Zg

[TRGNRD @TrigonRed]

@hybridherbst @sea3dformat @threejs @godotengine
How realistic is it to add something like this to Godot? It would be an absolute killer feature. AI-driven pipelines are becoming more and more relevant. We really need something like nanite for Godot.

[TRGNRD @TrigonRed]

@HuffBucket @RedPandaKoala 😂😂😂

[TRGNRD @TrigonRed]

@toyxyz3 The new nodes are simply much slower than the old ones. They’re harder to read and overall much less convenient to use. I literally took my workflow that works perfectly fine with the regular nodes, and with the 2.0 nodes it becomes completely unusable.

[TRGNRD @TrigonRed]

@sleenyre These are great news.

Wishing the Krea team good luck - genuinely excited for what's coming.