<h1 id="version-3.0-basicadvanced-now-with-even-less-llm!">Version 3.0 - Basic/Advanced - Now With Even Less LLM!</h1><p>Version 3.0 is out! It's basically the same as the others with a few minor QoL adjustments and (most importantly) it's tailored for my new <a target="_blank" rel="ugc" href="https://civarchive.com/models/3038852/q21-angle-lora-character-creation">Angle Lora for Q21.</a></p><p>Instructions are in the work flow, but if you've used the previous version you know how to use this one. It still uses my <a target="_blank" rel="ugc" href="https://github.com/AirbagGuy/ComfyUI-AirbagGuyAutoTrim">Auto Trim Node</a> as well.</p><p>Basically the goal of this version was consistency, which the new lora allows for. The prompts were trimmed down to utilize the lora's triggers instead of having to rely on Q21 to know what it should be each time. The Angle Lora itself has all of the angles shown in the Advanced version of the work flow. Honestly in theory you could forgo this work flow in general and just generate everything with a basic Q21 work flow if you'd prefer that (but that's no fun). You also don't even need to use the Angle Lora with this version either, as the triggers for the lora are intuitive prompts. Just know that it may not be consistent each time.</p><p>Any questions, issues, or idea please let me know!</p><p>Also, you're gonna want to mention footwear (or lack thereof) in your prompts.</p><h3 id="note-on-advanced-3.0-batched:">Note on Advanced 3.0 Batched:</h3><p>I made significant changes to the Advanced work flow. It is functions the same, but it forgoes individual samplers and instead goes with a batched approach. There's advantages and disadvantages to this. It's far more rigid and you will have to redo an entire run if an angle fails. However, it is far faster and because of the lora the run will only rarely fail.</p><p></p><h3 id="note-on-advanced-unbatched:">Note on Advanced Unbatched: </h3><p>This one functions closer to the original except a bit cleaner. There are now individual ksamplers, but they are all pulling from the same prompt list instead of individual prompts. Seeds you can find on the subgraph. Adjust those modify singular images. </p><p></p><p>**UPDATE: Quick note if you're using an older version of Version 3.0 (v3.1 to be specific), something I didn't realize when I posted this. "Head Notes" does not give any information to the Primary Heed. Those are all full body images, So the main node that's going to be doing the work is "Body Notes". That's the one you're going to use to build your feed images. In v3.2 this issue is fixed and Body Notes is renamed to "Feed Gen Notes". Behind notes still effects both. "Main Gen Only Notes" effects the Main feed. Head Notes and Behind will still their own separate thing. v3.1 will also allow you to add additional images to Main Gen.</p><h1 id="version-2.3-ph3-basic">Version 2.3 - pH3 Basic</h1><p>Obviously you need the Pruna H3 Image models to run it. Otherwise is almost purely a reference node swap out of the Q21 Version 2.1 Basic, with some minor tweaks:</p><p></p><p>First an additional section was put in for behind views. I actually have a Q21 version with that feature that I may upload later. Second, I added in a background removal nodes since (at least as far as I know) pH3 won't do alpha channels. I've also taken away the ability to toggle alpha channel on or off in the autotrim subgraph. It's still there if you want to bypass the background remover and would rather trust H3 to give you a white background. In my experience it often will put a shadow under the character, which will make it so the autotrim does not work properly (it works by measuring the differences between pixels).</p><p></p><p>I also tacked on refmods at the end if you want to just feed your images straight into it. It probably does not have the right VAE fed into it, but that's an easy fix. I did not actually try it. This was a bit of a rush job just because I wanted to see how the model functions.</p><h1 id="version-2.1-basicadvanced">Version 2.1 - Basic/Advanced</h1><p>***Update 10/09/2026 - Fixed FB Primary so it receives HS Primary as its first image.</p><p></p><p>***Note - After some experimenting I can confirm that this works with FK9 as well.</p><p></p><p>Excuse the mess. this is the advanced version of my Q21 Character Sheet Builder. It requires my AutoTrim node linked below:</p><p><a target="_blank" rel="ugc" href="https://github.com/AirbagGuy/ComfyUI-AirbagGuyAutoTrim">https://github.com/AirbagGuy/ComfyUI-AirbagGuyAutoTrim</a></p><p>Basically what it does is take One to Four inputs and turns it into various character images and character sheets. It is far more involved than the previous version, and of course takes WAY longer, but if you're patient you'll essentially get a whole dataset out of it. From there you can use these for reference images, refmods, or the basis of a dataset (See my Character Dataset Builder work flow - <a target="_blank" rel="ugc" href="https://civarchive.com/models/2719177/character-dataset-builder">https://civarchive.com/models/2719177/character-dataset-builder</a>). This is LLM free, there are a series of prompts built by Claude for the various angles.</p><p></p><h2 id="basic-vs-advanced">Basic vs Advanced</h2><p>Basic is actually a paired down, slightly adjusted version of Advanced, so it's going to work the same way. The differences?</p><p>-Advanced - Gives you 14 separate images plus 4 stitched character sheets (including Basic). There is far more angle diversity with those 14 images and tends to be more accurate. The downsides? Since there are a lot of chained generations you may see some (minor) quality degradation and color muddling. It also takes a long time since it's generating nearly four times as many images as Basic.</p><p>-Basic - Four views, simplified controls, and less laggy in comfy. It is, of course, way faster. It does however, in my opinion, create less accurate characters than Advanced.</p><p></p><p>My recommendation? If you have the patience use advanced. Advanced is going to give you everything Basic gets and more. If you're overwhelmed then try to learn basic first.</p><h2 id="directions:">Directions:</h2><h3 id="step-1-reference-images">Step 1 - Reference Images</h3><p>This takes 1 to 4 images. Upload and image to the Primary Image node, this one will be doing the heavy lifting. If you use more you'll have to manually connect them to the routing subgraph "Supplemental Reference Connection". One headshot will work; you'll just need to describe everything else.</p><h3 id="step-2-prompt">Step 2 - Prompt</h3><p>First thing you'll want to do is adjust "Primary Run Input Count" on the Prompt Templates subgraph to correspond with your number of reference images.</p><p>This works is that there are a series of pre-loaded prompts, which with bracketed keywords that correspond with the text nodes to the right. You fill out those text boxes, and a find/replace string node will replace the keywords and fill out the prompt.</p><p>-Style - Straight forward, dictate your desired style.</p><p>-Expression - Same deal, dictate your expression.</p><p>-Head/Body Changes - This section is used for explicit changes you want to make to your character, such as adjusting their outfit or appearance.</p><p>-Head/Body Notes - Used mostly for "Maintaining" features or calling things out explicitly you want to keep.</p><p>The prompt template will also let you choose if you want to use alpha channel or a background. I'd recommend alpha channel as you can replace the background later in the work flow.</p><h3 id="step-3-parameters">Step 3 - Parameters</h3><p>Three sections here, the model loader, and both size and seed parameters subgraphs. IF you've gotten this far I assume you know what everything on the Model Load node does. I did add a turbo toggle. This does work fairly well with turbo, I'd recommend viggle, you may need to adjust the manual sigmas though.</p><p>Parameters...Mess around with them and see what works best. Your head shots will have separate parameters from your body shots, as will your primary gen versus everything else. \</p><h3 id="step-4-stitch">Step 4 - Stitch</h3><p>Fairly straight forward. Everything uses my autotrim mode after everything is generated. The basic settings are on the subgraph. If you generate with a background you'll want to switch it off alpha channel. If you generate with alpha channel but want to add a background, turn "Flatten Images" on.</p><h3 id="step-5-save">Step 5 - Save</h3><p>Uses KJ's save nodes. Save categories are built in so all you need to do is add the save path and add what you want the file names to start with. It will also save the combined prompts as text output (I did not want to do it individually cuz...man...I'm tired enough)</p><h3 id="step-6-run">Step 6 - Run</h3><p>You all get this last part.</p><p></p><h2 id="version-2-updates:">Version 2 Updates:</h2><p>Now includes/requires my AutoTrim node</p><p><a target="_blank" rel="ugc" href="https://github.com/AirbagGuy/ComfyUI-AirbagGuyAutoTrim">https://github.com/AirbagGuy/ComfyUI-AirbagGuyAutoTrim</a></p><p>Also has a few more QoL features:</p><p>1) Input toggles for the first run. That's incase you want to use multiple references. It will choose a different starting prompt depending on the number of inputs. You'll still need to manually connect any additional images. It tops out at 4.</p><p>2) Toggle for alpha channel/white background</p><p>3) Improved model loader/parameters subgraph</p><p>4) Turbo Mode toggle - Can be used with your preferred turbo model, but you may need to update the manual sigmas.</p><p>In addition, I reworked the stock prompts so that the headshot will be at a 3/4 angle, which seems to help with</p><h2 id="overview">Overview</h2><p>This work flow will take an input image and generate four separate reference images that it will then save and stitch together into a proper character sheet.</p><p></p><p>This is meant for generating what we'll call "utility images" for making refmods, using as references for H3 and Q21, and for building a lora dataset. If you want a more artistic character sheet NeuroContent made a pretty damn good one <a target="_blank" rel="ugc" href="https://civarchive.com/models/2960750/qwen-image-21-character-design-sheet-maker-workflow?modelVersionId=3354175">here. </a>They also made one that's more utilitarian like mine, so give that look as well.</p><h2 id="directions">Directions</h2><p>Step 1: Upload one reference image. It can be full body or a headshot. If it is a headshot you will need to describe the outfit.</p><p>Step 2: Adjust input mega pixels. 1 MP goes a long way and makes the work flow work pretty fast, but you can go higher or bypass it if you wish.</p><p>Step 3: Adjust output mega pixels. Same deal, 1 tends to be fine but you can go higher if you can.</p><p>Step 4: Prompting - The way this works is that there are fixed outputs that will get you the proper angle. They will all generate with an alpha background. You can go into "Prompt Templates" and adjust them if you want, but I know from trial and error that these work. I tried doing it with an LLM and didn't like the time it took or the output, simpler seems to be better in this case.</p><p>*Prepended Prompt - Put in here whatever you want. I had it here from when the work flow used multiple reference but I don't use it anymore. Honestly for the sake of simplicity if you want to use multiple images try stitching them together first.</p><p>*Head Append - Because this is an edit model and edit models are very literal, I broke out the append prompt for the headshot into its own separate prompt. This helps to keep it focused on the head. If you mix in items below the shoulder or chest it will zoom it out and not give you a proper head shot. Not a bad thing depending on what you want, but I like to have the focused headshot.</p><p>*Body Append - This is for the full body images. This is where you can put your below-the-shoulders items. I like to just copy the Head Append Prompt and then just add onto it.</p><p>Step 5: Set your file name and save destination. I prefer KJ's save nodes because you can save outside of the output folder.</p><p>Step 6: Run</p><p></p><h2 id="other-notes">Other Notes</h2><p></p><p>YOU MUST HAVE SEPERATE SEEDS FOR EACH GENERATION! Q21 likes seed variation, it will fry images that have the same seed.</p><p>Also I have the "crop factor" set at 2. This helps to condense the output a bit more but you may not need and you may want to adjust if you change the full body generation resolutions.</p><p>Otherwise I tried to keep as much open to adjustment as you all might want. The settings as they stand are solid, but you change aspect ratios, output resolution samplers, and schedulers as much as you want.</p><p>Let me know if you find better settings!</p><p></p><p><br />Update: If you downloaded the work flying within 10 minutes of it being uploaded, there is an error. In the Save Cave subgraph, the "Stitch" text concatenate node, there is an extra line there that you will need to remove in order to automatically save the stitched image. It has since been fixed.</p>

Qwen 2.1 Character Sheet Builder - No LLM

Q21 Character Sheet Builder 3.1 - Advanced - Lora .json

Batched - Q21 Character Sheet Builder 3.2 - Advanced - Lora.json

Unbatched - Q21 Character Sheet Builder 3.3 - Advanced - Lora.json

Q21_CharSheet_V3.0_Adv

Q21_CharSheet_V3.0_Basic

pH3_CharSheet_v2.4_Basic

Q21_CharSheet_V2.1_Basic

Q21_CharSheet_V2.1_Adv

Q21_CharSheet_V2_Basic

Q21_CharSheet_V1

Files with SHA256: 0388a5351d5f5ed500e6b45116853c2705e4b34101757e7eb70df792658c0574