How can a 6B Model Outperform Larger Models in Photorealism!!!

https://redd.it/1pyr9ih
@rStableDiffusion

How are people combining Stable Diffusion with conversational workflows?

I’ve seen more discussions lately about pairing Stable Diffusion with text-based systems, like using an AI chatbot to help refine prompts, styles, or iteration logic before image generation.
For those experimenting with this kind of setup:
Do you find conversational layers actually improve creative output, or is manual prompt tuning still better?
Interested in hearing practical experiences rather than tools or promotions

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

Fal has open-sourced Flux2 dev Turbo.

https://preview.redd.it/kzd4n7gs37ag1.png?width=903&format=png&auto=webp&s=ab81e3133b7d4a0922d6ebfe03296eb87b30dbeb

fal/FLUX.2-dev-Turbo · Hugging Face



https://redd.it/1pyvw0q
@rStableDiffusion

Do you think Z-Image Base release is coming soon? Recent README update looks interesting

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

FYI: You can train a Wan 2.2 LoRA with 16gb VRAM.

I've seen a lot of posts where people are doing initial image generation in Z-Image-Turbo and then animating it in Wan 2.2. If you're doing that solely because you prefer the aesthetics of Z-Image-Turbo, then carry on.

But for those who may be doing this out of perceived resource constraints, you may benefit from knowing that you can train LoRAs for Wan 2.2 in ostris/ai-toolkit with 16GB VRAM. Just start with the default 24GB config file and then add these parameters to your config under the model section:

layer_offloading: true
layer_offloading_text_encoder_percent: 0.6
layer_offloading_transformer_percent: 0.6

You can lower or raise the offloading percent to find what works for your setup. Of course, your batch size, gradient accumulation, and resolution all have to be reasonable as well (e.g., I did batch_size: 2, gradient_accumulation: 2, resolution: 512).

I've only tested two different LoRA runs for Wan 2.2, but so far it trains easier and, IMO, looks more natural than Z-Image-Turbo, which tends to look like it's trying to look realistic and gritty.

https://redd.it/1pz0w56
@rStableDiffusion

Qwen Image Edit 2511: Workflow for Preserving Identity & Facial Features When Using Reference Images

https://preview.redd.it/lxp8ttxre8ag1.png?width=3920&format=png&auto=webp&s=2f68e028710c494eb9a02b718696f29c8f44b4d2

Hey all,

By now many of you have experimented with the official Qwen Image Edit 2511 workflow and have run into the same issue I have: the reference image resizing inside the TextEncodeImageEditPlus node. One common workaround has been to bypass that resizing by VAE‑encoding the reference images and chaining the conditioning like:

Text Encoder → Ref Latent 1 (original) → Ref Latent 2 (ref) → Ref Latent 3 (ref)

However, when trying to transfer apparel/clothing from a reference image onto a base image, both the official workflow and the VAE‑bypass version tend to copy/paste the reference face onto the original image instead of preserving the original facial features.

I’ve been testing a different conditioning flow that has been giving me more consistent (though not perfect) results:

Text Encoder → Ref Latent 1 → Ref Latent 1 conditions Ref Latent 2 + Ref Latent 3 → combine all conditionings

From what I can tell by looking at the node code, Ref Latent 1 ends up containing conditioning from the original image and both reference images. My working theory is that re‑applying this conditioning onto the two reference latents strengthens the original image’s identity relative to the reference images.

The trade‑off is that reference identity becomes slightly weaker. For example, when transferring something like a pointed hat, the hat often “flops” instead of staying rigid—almost like gravity is being re‑applied.

I’m sure there’s a better way to preserve the base image’s identity and maintain strong reference conditioning, but I haven’t cracked it yet. I’ve also tried separately text‑encoding each image and combining them so Ref Latent 1 isn’t overloaded, but that produced some very strange outputs.

Still, I think this approach might be a step in the right direction, and maybe someone here can refine it further.

If you want to try the workflow, you can download it here:
**Pastebin Link**

Also, sampler/scheduler choice seems to matter a lot. I’ve had great results with:

er\_sde (sampler)
bong_tangent (scheduler)

(Requires the **RES4LYF** node to use these with KSampler.)

https://redd.it/1pz2gxy
@rStableDiffusion

Wan 2.2 Motion Scale - Control the Speed and Time Scale in your Wan 2.2 Videos in ComfyUI
https://youtu.be/Zmkn6_vyMN8

https://redd.it/1pz2kvv
@rStableDiffusion