<h1 id="anima-regional-lora">Anima Regional LoRA</h1><p>while scrolling in X I saw some share their Regional LoRA techniques, many of those workflows were over complicated and uses outdated methods sometimes.</p><p>I remember when I was trying to do this 2 years ago it was not trivial. so I decided to upload multiple workflow examples avoiding custom nodes as much as possible.</p><p>there are many good Regional LoRA workflows out there and I am sure they are way better than mine. but mine is meant to be simple, understandable and uses native comfyui nodes.</p><p></p><h1 id="what-you-need">What You Need</h1><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426/anima?modelVersionId=2945208">Anima base</a> can be replaced with your desired model, even Illustrious XL if you know what you are doing.</p><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/2466415/cosmos-predict25-2b-base-distilled-extracted-dmd2-lora">Speed LoRA</a> can also be skipped but honestly you should use speed LoRA even with high end pc, the speed is too good to pass and this one even improve quality over base.</p><p>and for SAM3 example you need the Comfy-Org <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/sam3.1">SAM3.1 model</a></p><p>and finally for the upscale example <a target="_blank" rel="ugc" href="https://github.com/ssitu/ComfyUI_UltimateSDUpscale"><strong>ComfyUI_UltimateSDUpscale.</strong></a></p><p>and all these workflows were built and tested specifically for this <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.24.0">comfyui</a> version.</p><p><strong>⚠️ Tip:</strong> Add <code>--disable-dynamic-vram</code> to your ComfyUI launch flags. On my RTX 4090, without it generation is about 50% slower. Your mileage may vary but it's worth trying because you will need generation Speed.</p><p></p><h1 id="the-method:-lora-hook">The Method: LoRA Hook</h1><p>The main method used across all workflows is ComfyUI's native <a target="_blank" rel="ugc" href="https://blog.comfy.org/p/masking-and-scheduling-lora-and-model-weights">LoRA Hook</a> node.</p><p>Think of it like Regional Conditioning, but instead of only separating conditions (prompt) it also applies a separate LoRA to each region. For simplicity, assume the KSampler generates one image for each regional condition before giving the final image. As a general rule I advise having one region per character plus one extra for the background — so 3 regions for 2 characters, 4 for 3 characters, and so on.</p><p>Yes, this multiplies generation time with each region added. But in my testing this is by far the best method for regional LoRA, especially when characters overlap. The quality difference is worth it and on top of that it works perfectly with ultimate sd upscale.</p><p>Beyond the conditioning setup, the real challenge is <strong>how you define the regions</strong> — and that's exactly why there are 4 workflows and 3 masking approaches here.</p><p></p><h2 id="create-hook-lora:">Create Hook LoRA:</h2><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/12a556d9-0388-4367-bd55-b07b259c6542/original=true/12a556d9-0388-4367-bd55-b07b259c6542.jpeg" /></p><p>remember to choose the LoRAs in the Create Hook LoRAs. the weights work slightly different from usual btw, so you might need to go slightly up or down.</p><p></p><h2 id="prompting-advice:">Prompting Advice:</h2><p>all workflows use 3 prompt sections</p><p><strong>1- General prompts: </strong>here you write the prompts for the style, background and clothing and action if they are the same between the 2 character</p><p>2- <strong>girl left:</strong> here you write the activation tags for left Character LoRA and clothing, action and body type if they are not the same between characters.</p><p>3- <strong>girl right:</strong> here you write the activation tags for right Character LoRA and clothing, action and body type if they are not the same between characters.</p><h3 id="example">example</h3><p>General prompts: masterpiece, 2girls. side-by-side, shoulder to shoulder, outdoors, grass field, sunset, summer dress, sun hat, spaghetti strap, white dress, full body,</p><p>girl left: [character tags], looking back, light smile, from behind, medium breasts, cleavage</p><p>girl right: [character tags], looking to the side, seductive smile, from side, small breasts,</p><p>this will produce results like this</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/edb8e7e4-bc3d-4301-84e6-89506acda5ca/original=true/edb8e7e4-bc3d-4301-84e6-89506acda5ca.jpeg" /></p><h2></h2><h1 id="workflow-breakdown">Workflow Breakdown</h1><h2 id="1-regionalloraanima5050">1- RegionalLoraAnima5050</h2><p>The simplest entry point. The canvas is automatically split 50/50 left and right — no mask drawing, no external files. Two EmptyImage nodes create a pure black and white image, stitched together and converted to masks dynamically. Resolution-aware, fully native, zero custom nodes. Just load your models, write two prompts, and generate. Best for symmetrical side-by-side compositions.</p><p>this method will only produce 2 regional conditions so no background condition.</p><h2 id="2-regionalloraanimareference">2- RegionalLoraAnimaReference</h2><p>Want more control over character placement? Load your reference image in the LoadImage nodes, right-click to open the mask editor, and draw around each character manually. Then the workflow will merge the 2 masks and invert them for the background mask. this will result in 3 regional conditions.</p><p>I advise not drawing precise mask and give more breathing room to each character.</p><p><strong>NOTE: </strong>The masks won't force the model to generate inside them, the model can generate the characters despite the regional condition, if you want bizarre poses try to use controlnet as well.</p><p>I have provided the reference image I used for the example and even the reference image with the mask as alpha region, if you want to test load the LeftMask.png in Load Image (as Mask) - Left node and the RightMask.png in Load Image (as Mask) - Right then make sure you connect the mask output of each node to the appropriate reroute node.</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/04deac35-3128-47b7-8f12-ec47b4026120/original=true/04deac35-3128-47b7-8f12-ec47b4026120.jpeg" /></p><h2 id="3-regionalloraanimasam3">3- RegionalLoraAnimaSam3</h2><p>Generate Draft image first using the General prompt. and from the draft we use Sam3.1 to generate mask for each character. Each mask is grown by 64px for breathing room, then the other character's original mask is subtracted to prevent overlap. and then we use this mask for the regional condition.</p><h2 id="4-upscaleregionalloraanimasam3">4- UpscaleRegionalLoraAnimaSam3</h2><p>This is the SAM3 workflow with upscaling added on top. After generating the final image with both character LoRAs applied, the workflow re-runs SAM3.1 to re-mask the characters — because the composition likely shifted from the draft. Fresh masks are then used to create regional conditions again before feeding into Ultimate SD Upscale. This workflow is SLOW. Really SLOW. But honestly, worth it to see your waifus in 4K.</p><p></p><h1 id="notes:">Notes:</h1><p>These workflows are basic. use your logical mind to utilize those techniques in more advanced workflows.</p><p>Those Workflows were tested in <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.24.0">comfyui release v0.24.0. </a>so I apologize if they don't work in older versions. the workflows were tested in RTX 4090, RTX 5070 Ti and rx 7900 xtx. I am not sure if they will work in 8gb vram. if you tested and it work please tell me!</p><p></p><p>If you want to add more LoRA to each region (each region uses a character LoRA and style LoRA as well) simply connect another <strong>create LoRA hook</strong> node in serious like the following image.</p><p><img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/ee451d48-3962-4ceb-9d48-e928c4177348/original=true/ee451d48-3962-4ceb-9d48-e928c4177348.jpeg" /></p>

Files with SHA256: a5a943021d2e02b32f75475a46cc7fcd742d9c9b0843f7c1c7cf524fcda0104e