<p><span style="color:rgb(250, 82, 82)">By downloading this model, you agree not to use it for any commercial purpose, including selling images generated using this model. If you are interested in commercial licensing or commissioning a model, please contact me and we can discuss your use case.</span></p><p><span style="color:rgb(253, 126, 20)"><em>If you use my models, please take the time to post some images to the pages so I can see how they are being used and make improvements in the future. Also feel free to reach out if you have ideas for future models!</em></span></p><h2 id="overview">Overview</h2><p>This model is my first attempt at a multi-concept amputee model, and it really shows off what Qwen-2.5 VL is capable of as a text encoder. I tried for a long time to accomplish something like this with Flux and was not able to do so, mainly because t5xxl couldn't represent the concepts with enough fidelity.</p><h2 id="how-to-prompt-(please-read)">How to Prompt (Please Read)</h2><p>Qwen strongly prefers prose based prompts. This applies to negative prompts as well.</p><p>This model does not utilize a trigger word, and in general, I would recommend against trigger word approaches with Qwen, as its text encoder is significantly more expressive than what we have seen from older generation models based on CLIP or t5xxl.</p><p>Describe limb differences in a semi-clinical fashion. "She has bilateral hip disarticulations" works much better than "She is a dhd amputee" since common abbreviations like "dhd", "dae", etc don't encode to vectors with reasonable relationships during text encoding but the text encoder absolutely understands hips, shoulders, and disarticulation as English words.</p><p>Plan to spend a decent amount of words and sentences on the limb schema. Generally each caption contains 1-3 sentences of information, and sometimes more for complex/inconsistent schemas.</p><h3 id="limb-differences">Limb Differences</h3><p>Here are some examples of how to describe some limb differences:</p><ul><li><p>She has bilateral {hip|shoulder} disarticulations</p></li><li><p>She has a {left|right} {above|below} knee leg amputation</p></li><li><p>She has a right {hand|foot} amputation</p></li><li><p>She is a limbless amputee. She has bilateral hip and shoulder disarticulations.</p></li></ul><p>You can try describing the length of residual limbs, or where they end. Generally the model gets it to a point, but the more picky you are the lower the hit rate.</p><h3 id="fixing-extraneous-amputations">Fixing Extraneous Amputations</h3><p>The model sometimes removes more limbs than you asked for, especially when you request asymmetric schemas like ooe, since like most models, Qwen Image expects lateral symmetry in human bodies.</p><p>If you find this happening, the fix is pretty simple: describe the intact limbs clearly. Where are they? Is the elbow bent? Etc. Add details about them until they are showing up reliably. With a good prompt, I can usually get an 80-90% hit rate on asymmetric schemas.</p><h3 id="lefts-and-rights">Lefts and Rights</h3><p>Unfortunately, Qwen is not good at lefts and rights, and this form of fine tuning is not able to repair that across the board, despite careful captioning and consistent use of stage directions, which is the best practice. This likely won't be fixable with Qwen Image. You can always flip the image horizontally after the fact if you're really picky about it.</p><h3 id="scarring">Scarring</h3><p>This model has much more information about scars in its training set than my Flux models. The captions used during training took this into account. Here's some captioning advice:</p><ul><li><p>For disarticulations, refer to the scarred areas as "amputation sites".</p></li><li><p>For limb or hand amputations, refer to the scarred areas as "residual limbs"</p></li></ul><p>This was done consistently, and helps reinforce the differences between disarticulations and amputations with a residual limb.</p><p>Sample phrasing:</p><ul><li><p>She has bilateral shoulder disarticulations with faint scarring at her shoulder amputation sites</p></li><li><p>She has bilateral hip and shoulder disarticulations with faint scarring at her amputation sites</p></li><li><p>She has bilateral above-knee amputations with short residual limbs. She has bilateral above-elbow amputations with very short residual limbs. She is supported by distal ends of her residual arms and legs.</p></li><li><p>Her amputation sites are smooth.</p></li><li><p>Her residual limbs are smooth.</p></li></ul><p>I recommend avoiding the term "linear scars" as it tends to produce weird behavior. "scar tissue" is OK and will produce slightly different results than "scarring" which you may prefer. If the scars are coming out too gorey or wound-like, mention "wounds" in your negative prompt.</p><p>Finally, some images in the training set were described as "Electronic Surgery Images" if they were built in photoshop. I have not experimented much with prompting for this, but it's something you can play with if you have some nostalgia for that vibe.</p><h3 id="photorealism">Photorealism</h3><p>Qwen's photorealism and lighting choices are not very good without prompting for specifics. You'll want to explicitly caption for photographic qualities that you find desirable:</p><ul><li><p>crisp 35mm photograph of .... Subtle film grain enhances analog texture. The lighting is cool, even, diffused</p></li><li><p>35mm photograph of ... The gallery's recessed spotlights cast dramatic light across her torso and shoulders, emphasizing ... . Subtle film grain enhances analog texture while...</p></li><li><p>35mm Photograph of ... The setting features even, diffused daylight highlighting her complexion, with soft shadows emphasizing the contrast of...</p></li></ul><h3 id="text">Text</h3><p>Qwen is great at text, and you can go wild with tattoos, captions, overlays, etc in a way that was not possible with Flux. However, be aware that text can leak unintended impacts onto the rest of the image.</p><p>If you are playing with text, try the same prompt/seed with and without the text to understand the tradeoffs. If you don't like what happens, you can always use Qwen Edit to add the text after the fact if it's causing you issues.</p><h2 id="about-the-sample-images">About the Sample Images</h2><p>All sample images use the simplest Qwen ComfyUI workflows, with Qwen Image fp8 as the base, and no other LoRAs loaded. 20 Steps, Euler, Guidance of 3.0-4.0, and lora weight=1.0.</p><h2 id="training-details">Training Details</h2><p>This model was trained with a new, 500-image dataset that was LLM-captioned using Gemma3 12B, and then all of the concept-specific aspects were captioned by hand, one at a time.</p><p>I also used a data set of 10,000 high quality photographs for regularization which helps the model generalize without overfitting while fixing some of Qwen's built-in tropes and defects. The regularization dataset was also captioned using Gemma3 12B.</p><p>Generating NSFW content was not a goal during model training, and NSFW images will exhibit typical Qwen Image defects. Feel free to experiment with other LoRA's if that is your goal. In my experience so far, mixing Loras works much better with Qwen Image than it did with Flux.</p><h2 id="commercial-use-policy">Commercial Use Policy</h2><p>By downloading this model, you agree not to use it for any commercial purpose, including selling images generated using this model. If you are interested in commercial licensing or commissioning a model, please contact me and we can discuss your use case.</p>

Files with SHA256: 39c38ce257d406cb1be2d7e30d4bf3d2636ada8f7e17b72a59a450b3dc9b47d2