Сonsistency characters V0.4 | Generate characters only by image and prompt, without character's Lora! | IL\NoobAI Edit

https://redd.it/1okrsld
@rStableDiffusion

Update — FP4 Infrastructure Verified (Oct 31 2025)

Quick follow-up to my previous post about running SageAttention 3 on an RTX 5080 (Blackwell) under WSL2 + CUDA 13.0 + PyTorch 2.10 nightly.

After digging into the internal API, I confirmed that the hidden FP4 quantization hooks (scaleandquantfp4, enableblockscaledfp4attn, etc.) are fully implemented at the Python level — even though the low-level CUDA kernels are not yet active.

I built an experimental FP4 quantization layer and integrated it directly into nodesmodelloading.py.
The system initializes correctly, executes under Blackwell, and logs tensor output + VRAM profile with FP4 hooks active.
However, true FP4 compute isn’t yet functional, as the CUDA backend still defaults to FP8/FP16 paths.


---

Proof of Execution

attention mode override: sageattn3
FP4 quantization applied to transformer
FP4 API fallback to BF16/FP8 pipeline
Max allocated memory: 9.95 GB
Prompt executed in 341.08 seconds


---

Next Steps

Wait for full NV-FP4 exposure in future CUDA / PyTorch releases

Continue testing with non-quantized WAN 2.2 models

Publish an FP4-ready fork once reproducibility is verified


Full build logs and technical details are on GitHub:
Repository: github.com/k1n0F/sageattention3-blackwell-wsl2

https://redd.it/1oktwaz
@rStableDiffusion

Brie's Lazy Character Control Suite

https://redd.it/1ol0a4h
@rStableDiffusion

Raylight, Multi GPU Sampler. Finally covering the most popular models: DiT, Wan, Hunyuan Video, Qwen, Flux, Chroma, and Chroma Radiance.

https://redd.it/1okwh6o
@rStableDiffusion

Anyone else think Wan 2.2 keeps character consistency better than image models like Nano, Kontext or Qwen IE?

I've been using Wan 2.2 a lot the past week. I uploaded one of my human AI characters to Nano Banana to get different angles to her face to possibly make a LoRA.. Sometimes it was okay, other times the character's face had subtle differences and over time loses consistency.

However, when I put that same image into Wan 2.2 and tell it to make a video of said character looking in a different direction, its outputs look just right; way more natural and accurate than Nano Banana, Qwen Image Edit, or Flux Kontext.

So that raises the question: Why aren't they making Wan 2.2 into its own image editor? It seems to ace character consistency and higher resolution seems to offset drift.

I've noticed that Qwen Image Edit stabilizes a bit if you use a realism lora, but I haven't experimented long enough. In the meantime, I'm thinking of just using Wan to create my images for LoRAs and then upscale them.

Obviously there are limitations. Qwen is a lot easier to use out of the box. It's not perfect, but it's very useful. I don't know how to replicate that sort of thing in Wan, but I'm assuming I'd need something like VACE, which I still don't understand yet. (next on my list of things to learn)

Anyway, has anyone else noticed this?

https://redd.it/1ol2bsm
@rStableDiffusion