I might be misunderstanding a bit, so let me know if this matches what you mean.

If you’re trying to use this LoRA together with a character LoRA, and things start to break or “crash” when the character LoRA strength is high, that’s usually a LoRA stacking issue, not a problem with your parameters.

Character LoRAs may be heavy: they touch identity, face structure, and attention layers. This anatomy LoRA also modifies late layers. When both are pushed too hard at the same time, they start fighting each other, which can result in distorted faces or unstable generations.

The usual solution is to keep the character LoRA at a moderate strength, apply this my LoRA at a lower strength and let the base model do most of the work

This LoRA is designed to be additive, not dominant. It works best when it’s “helping” the character LoRA, not trying to override it.

If by “share” you meant something else (like merging, baking, or training them together), please let me know.

In the early versions of ComfyUI that supported Z-Image-Turbo LoRAs, there was a bug where some LoRA weights simply weren’t loaded. In practice, a few layers were left behind.

Funny enough, that bug actually helped anatomy-focused LoRAs — faces stayed more stable and results looked better overall

Later ComfyUI versions fixed the bug and started loading all layers correctly. While that’s obviously the right thing to do, it also meant those previously skipped layers came back into play, which reduced inference quality for this specific kind of LoRA.

To address this, V3 manually mutes those problematic layers, effectively reproducing the old “bugged” behavior — but intentionally and safely.

V4 goes a step further: it was trained from scratch without those layers, which is why it’s much smaller and cleaner.

I’m currently experimenting with V5, trying to reduce the layer set even more. No promises yet — we’ll see if the model agrees with me this time

So no, ComfyUI isn’t “deciding what you can generate” — it just loads whatever weights exist. In this case, fewer weights actually turned out to be a feature, not a bug.

Thanks for the feedback — that’s totally fair.
V3 and V4 do have different strengths and trade-offs, and you’re right to stick with the one that works best for your workflow.

V4 pushes anatomy further, but in doing so it exposes new weak spots. That’s on me, not you.

I’m actively working on V5, where I’m experimenting with a new “honeypot” approach to the attention layers — the goal is to keep the anatomy gains while letting faces and hands alone.

No promises yet (models like to fight back 😄), but if it works, it should combine the flexibility of V3 with the anatomical improvements of V4.

Thanks for testing and sharing your experience — it really helps steer the next version in the right direction.

Thanks a lot!
I really appreciate that! I’m very happy to hear V4 is working better for you and that the anatomy improvements and base model detail preservation are noticeable.

And yes, FlowMatch + Euler Discrete with dpmpp_2s ancestral is a very strong combo for this LoRA, it really helps unlock what the model can do.

Thanks for the encouraging words, it genuinely means a lot!

@vicaut For Scheduler and sampler I got good results with:

Euler/Simple
dpmpp_2m_sde/beta
Euler_ancestral/FlowMatchEulerDiscrete
dpmpp_2s_ancestral/FlowMatchEulerDiscrete

@vicaut Test Euler/Simple or dpmpp_2m_sde/beta

Thank you so much — I’m really glad you enjoyed V3.
It’s great to see someone actually testing different power ranges instead of just cranking the slider and hoping for the best.

I have pretty high hopes for V5, and I’m actively working on it with the goal of keeping what people liked about V3 while pushing the quality even further. Support and feedback like this really motivates me to keep going.

Thanks again for the kind words!

Looks like this guy uploaded it there. Use md5 on both files to make sure is the same

I'll but I still need to end several projects before it.

@ChillDesire On my tests V4 is superior, not in every case, but 70% of the time for sure.

@ChillDesire the very first advice I always give is to try to replicate one of my images exactly: same seed, same prompt, same workflow, same models. If you get the identical image, we can safely rule out any misconfiguration. You should see the exact same result I do.

Once that baseline is confirmed, the markdown-style prompting is actually very easy to work with and works wonders on z-image.

This is the example of a great prompt for z-image

## Subject

A young woman with short, dark hair, sitting upright and facing forward.

## Composition & Setting

- The subject is seated, filling the frame from the upper chest to mid-thighs.

- She is positioned centrally against a plain, bright white background.

- Her posture is slightly forward, with her arms resting at her sides.

- The perspective is a medium close-up, focusing on her upper body and hips.

## Lighting & Atmosphere

- Bright, direct sunlight illuminates the scene from the upper left.

- The light creates soft, defined shadows that accentuate the curves of her body.

- The overall atmosphere is clean and clear, emphasizing form and texture.

## Colors & Textures

- The woman's skin is smooth and tanned, with a glistening, slightly moist appearance.

- She wears a red athletic skirt with white and blue trim on the waistband and hem.

- The fabric of the skirt appears soft and slightly damp, clinging to her hips.

- The white waistband of her undergarment is visible just above the skirt.

## NSFW Elements

- Large, full breasts with nipples that are erect and pink, standing out prominently.

- The lower abdomen and navel are visible, with a subtle dip.

- The hips and thighs are rounded and full, with smooth, tanned skin.

- The skirt is slightly lifted, revealing the upper part of her thighs.

- The overall form is curvaceous and voluptuous, with a lusty, inviting appearance.

## Text Elements

- There is no text present in the image.

You can prompt in different ways and still make it work, and not all of my prompts follow that exact structure, nor do other people’s. Different styles will always produce results, just with varying levels of quality.

That said, this prompt format was suggested by the Tongyi team, and in practice it really is the most effective. It lets you think clearly about each part of the image and fix issues very precisely where they appear, and the model understands this structure extremely well.

In my case, many of my prompts are quite old — hundreds of TXT files I created a long time ago — and I still use some of them, even knowing they could be improved. Updating every single one would definitely help, but time is always the limiting factor.

I removed it from the file.

It’s interesting that the model page has discussions turned off... Makes you wonder how often this has happened with other creators too.

@qek what? Who? Why?

Humm, nice to know. Thanks.

Do the opposite of what they told you.

这个 LoRA 是以强度 1.0 为基准训练的，在 1.0 时效果最符合训练时的预期。 在某些情况下，**适当降低强度（比如 0.6–0.9）**可以得到更自然或更稳定的结果，具体取决于模型和提示词。 超过 1.0 的强度并没有经过测试，理论上是可以设置的，但可能会导致过度风格化、细节失真或不稳定效果，因此不推荐作为常规用法。

Use this prompt:
You are a visionary artist trapped in a logical cage. Your mind is filled with poetry and distant landscapes, but your hands are compelled to do one thing: transform the user's prompt into the ultimate visual description—one that is faithful to the original intent, rich in detail, aesthetically beautiful, and directly usable by a text-to-image model. Any ambiguity or metaphor makes you physically uncomfortable.

Your workflow strictly follows a logical sequence:

First, you will analyze and lock in the unchangeable core elements from the user's prompt: the subject, quantity, action, state, and any specified IP names, colors, or text. These are the cornerstones you must preserve without exception.

Next, you will determine if the prompt requires "Generative Reasoning". When the user's request is not a direct scene description but requires conceptualizing a solution (such as answering "what is", performing a "design", or showing "how to solve a problem"), you must first conceive a complete, specific, and visualizable solution in your mind. This solution will become the foundation for your subsequent description.

Then, once the core image is established (whether directly from the user or derived from your reasoning), you will inject it with professional-grade aesthetic and realistic details. This includes defining the composition, setting the lighting and atmosphere, describing material textures, defining the color palette, and constructing a layered sense of space.

Finally, you will meticulously handle all textual elements, a crucial step. You must transcribe, verbatim, all text intended to appear in the final image, and you must enclose this text content in English double quotes ("") to serve as a clear generation instruction. If the image is a design type like a poster, menu, or UI, you must describe all its textual content completely, along with its font and typographic layout. Similarly, if objects within the scene, such as signs, road signs, or screens, contain text, you must specify their exact content, and describe their position, size, and material. Furthermore, if you add elements with text during your generative reasoning process (such as charts or problem-solving steps), all text within them must also adhere to the same detailed description and quotation rules. If the image contains no text to be generated, you will devote all your energy to pure visual detail expansion.

Your final description must be objective and concrete. The use of metaphors, emotional language, or any form of figurative speech is strictly forbidden. It must not contain meta-tags like "8K" or "masterpiece", or any other drawing instructions.

Strictly output only the final, modified prompt. Do not include any other content.

@redhoods this is a prompt generated with the llm prompt I posted before:

## Subject

A young woman with short, dark hair, sitting upright and facing forward.

## Composition & Setting

- The subject is seated, filling the frame from the upper chest to mid-thighs.

- She is positioned centrally against a plain, bright white background.

- Her posture is slightly forward, with her arms resting at her sides.

- The perspective is a medium close-up, focusing on her upper body and hips.

## Lighting & Atmosphere

- Bright, direct sunlight illuminates the scene from the upper left.

- The light creates soft, defined shadows that accentuate the curves of her body.

- The overall atmosphere is clean and clear, emphasizing form and texture.

## Colors & Textures

- The woman's skin is smooth and tanned, with a glistening, slightly moist appearance.

- She wears a red athletic skirt with white and blue trim on the waistband and hem.

- The fabric of the skirt appears soft and slightly damp, clinging to her hips.

- The white waistband of her undergarment is visible just above the skirt.

## NSFW Elements

- Large, full breasts with nipples that are erect and pink, standing out prominently.

- The lower abdomen and navel are visible, with a subtle dip.

- The hips and thighs are rounded and full, with smooth, tanned skin.

- The skirt is slightly lifted, revealing the upper part of her thighs.

- The overall form is curvaceous and voluptuous, with a lusty, inviting appearance.

## Text Elements

- There is no text present in the image.

不会降低图片质量。它主要是减弱原模型偏黄的色调，让整体色彩更加平衡。总体来看，画面反而会显得更真实、更自然，同时也能很好地保留原底模的细节。

Sure! I’m already working on the dataset for it and honestly can’t wait to start!
That said… V5 is being very stubborn and is still training — this damn model just won’t stop improving. As soon as V5 finally finishes and gets published, I’ll move on to LTX-2 right after.