huehuehue9001 Posted February 16 Posted February 16 (edited) I heard you like realistic. Spoiler workflow is nova asian > nova asian 2x upscale > qwen 4k upscale > uncanny valley face inpainting > nova asian nipples inpainting. the pubes and cunny is qwen. the areola is qwen, but the nipple is very plastic, so i had to inpaint. qwen have better realistic skin than illustrious, sdxl or noob. why not just start with qwen? qwen keeps giving the same pose and camera angle all over again. uncanny valley checkpoint gives the most similar face to the game and with realistic lighting and skin. Spoiler Nanami's face wouldn't come up looking like the game. Edited February 17 by huehuehue9001 5
5250 Posted February 19 Posted February 19 (edited) Pos night,(1girl,solo,alone,solo focus,looking at viewer:1.4),embarrassed,cowboy shot,thighs,<lora:IL_DOAXVV:0.75>,__DOAXVV__,<lora:realistic filter [IL]:0.7>, lying on her stomach on a bed,ass focus,viewed from behind,pubic hair,, masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed,realistic, absurdres, scenery, photorealistic, depth of field, photorealistic details,realistic lighting,Cinematic composition,high contrast,moody and suspenseful ambiance,while the rest of the space fades into black, Neg anime, 3d, rendered, modern, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra digits, fewer digits, very displeasing, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, signature, watermark, username, simple background, conjoined, bad ai-generated Edited February 19 by 5250 8
5250 Posted February 19 Posted February 19 (edited) If the initial resolution exceeds 1700Pics, the body becomes vertically elongated and the image often looks strange. I wonder if there is a way to output an image with 4K level resolution in a single rendering without using an upscaler. Edited February 19 by 5250
huehuehue9001 Posted February 19 Posted February 19 (edited) You have to upscale. SDXL, noob, IL is trained at 1 megapixel. You can force it above, but expect errors. I used to do a batch of 1600x900, but then I gave up because I'm going to upscale it anyway and concentrate on getting a good composition for base image. I also get almost no errors for base 1MP image. Just the lack of detail face and green thing, which is easily fixable. You don't have to adetailer, the upscale will fix it. Elongated limbs or torso is hard to fix. so it's better to restart. Edited February 20 by huehuehue9001 1
5250 Posted February 20 Posted February 20 15 hours ago, huehuehue9001 said: You have to upscale. SDXL, noob, IL is trained at 1 megapixel. You can force it above, but expect errors. I used to do a batch of 1600x900, but then I gave up because I'm going to upscale it anyway and concentrate on getting a good composition for base image. I also get almost no errors. Thanks for your advice.
5250 Posted February 21 Posted February 21 (edited) On 2/19/2026 at 9:22 PM, huehuehue9001 said: You have to upscale. SDXL, noob, IL is trained at 1 megapixel. You can force it above, but expect errors. I used to do a batch of 1600x900, but then I gave up because I'm going to upscale it anyway and concentrate on getting a good composition for base image. I also get almost no errors for base 1MP image. Just the lack of detail face and green thing, which is easily fixable. You don't have to adetailer, the upscale will fix it. Elongated limbs or torso is hard to fix. so it's better to restart. Hello. I didn't want to use Upscaler because of need more time to generate. But I tried to use Upscaler follow your advice. I am surprised!! I was under the impression that using Upscaler would take a long time........ But, The image generation time was only about 1 second different when using Upscaler. Therefore, from now, I decided to always use Upscaler. Thank you for your advice. The purpose of the test is to check the image quality and time required to generate W1280*H2048 pics image. Test environment. A : Use Upscaler Sampling step : 20 >>> 15 seconds Sampling step : 25 >>> 19 seconds Sampling step : 30 >>> 21 seconds B : Not use Upscaler Sampling step : 20 >>> 14 seconds Sampling step : 25 >>> 18 seconds Sampling step : 30 >>> 21 seconds Edited February 22 by 5250 2
5250 Posted February 22 Posted February 22 (edited) When generating using Upscaler, it seemed to be slower, so after the previous test, which showed almost no difference when generating one image, I tried testing the usage time when generating 4 and 8 images simultaneously. As expected, we found that when the number of simultaneous generations was increased, using Upscaler became significantly slower. The correct answer is to follow the basic method of generating a large number of images simultaneously without Upscaler, and then regenerating only the good images with Upscaler. (Maybe Z-image is more fast, but I'll leave it for a while because I'd have to rebuild Lora and there aren't many compatible Lora's yet.) A.Use Upscaler + Batch size 4 Sampling step : 20 >>> 63 seconds Sampling step : 25 >>> 74 seconds Sampling step : 30 >>> 88 seconds B.Use Upscaler + Batch size 8 Sampling step : 20 >>> 124 seconds C.Not use Upscaler Sampling step : 20 >>> 49 seconds Sampling step : 25 >>> 61 seconds Sampling step : 30 >>> 70 seconds D.Use Upscaler + Batch size 8 Sampling step : 20 >>> 99 seconds Edited February 22 by 5250
5250 Posted February 22 Posted February 22 (edited) 4K pics use Upscaler. (Generated 6 images not use upscaler, then the best pics use upscaler. this way was the fastest method.) Edited February 22 by 5250 11
huehuehue9001 Posted February 22 Posted February 22 (edited) Always try new things. Especially if it's already ready setup. The body looks nicely proportionate now but the bicep is a bit long on some. Look at the elbow. Try lowering denoise to 0.25 to 0.4. As for resolution, quoted from comfyui template. Quote NOTE: For SDXL, it is recommended to use trained values listed below: - 1024 x 1024 - 1152 x 896 - 896 x 1152 - 1216 x 832 - 832 x 1216 - 1344 x 768 - 768 x 1344 - 1536 x 640 - 640 x 1536 A safe upscale is 1.5x upscale with these resolutions. But you can go 2x or lower. Just look for deformity. You can try lowering the denoise and/or cfg. Or fix with inpainting manually. The simple workflow of hiresfix is, t2i ksampler > upscale image > ksampler again. So it's really just img2img. You can use lanczos. I tested RealESRGAN and lanczos, it's almost the same. Do test it yourself. If you press the button on the right. It will do the hiresfix step of the last image generated. So you can t2i base image without hiresfix on > change prompt if you want. > press orange button. If you prompt "looking down" you can change the eyes to look down but not the head tilting down. It took me awhile to find that button lol. As for me, I do it all in comfyui now. I uploaded my comfy json. I don't use adetailer anymore. I just manually inpaint at the end. sdxl_t2i.jsonsdxl_hiresfix_2steps.jsonInpainting.json Edited February 22 by huehuehue9001 1
5250 Posted February 22 Posted February 22 (edited) 17 minutes ago, huehuehue9001 said: Always try new things. Especially if it's already ready setup. The body looks nicely proportionate now but the bicep is a bit long on some. Look at the elbow. Try lowering denoise to 0.25 to 0.4. As for resolution, quoted from comfyui template. A safe upscale is 1.5x upscale with these resolutions. But you can go 2x or lower. Just look for deformity. You can try lowering the denoise and/or cfg. Or fix with inpainting manually. The simple workflow of hiresfix is, t2i ksampler > upscale image > ksampler again. So it's really just img2img. You can use lanczos. I tested RealESRGAN and lanczos, it's almost the same. Do test it yourself. If you press the button on the right. It will do the hiresfix step of the last image generated. So you can t2i base image without hiresfix on > change prompt if you want. > press orange button. It took me awhile to find that button lol. As for me, I do it all in comfyui now. Hello. Thank you for your advice! I actually want to use comfyUI, not Forge. I usually use ComfyUI to make Video or music etc. Only t2i still use Forge. I can't make best workflow for t2i on ComfyUI...... I use same Checkpoint & same Lora and same etc, but not same result between ComfyUI and Forge. I will try to use ComfyUI again.... Thank you. Edited February 22 by 5250
5250 Posted February 22 Posted February 22 (edited) Why are Forge and ComfyUI so different under the same conditions...... A.Forge B.ComfyUI Edited February 22 by 5250
huehuehue9001 Posted February 22 Posted February 22 (edited) I don't have such UI to toggle characters. I have a notepad. lol. The comfy generation face looks bad. It's also blurry. You can drag and drop forge image into comfy ui and see the forge workflow. tsukushi nova asian 1344x768 t2i. 1440p hiresfix upscale. Edited February 22 by huehuehue9001 8
5250 Posted February 23 Posted February 23 (edited) On 2/23/2026 at 12:29 AM, huehuehue9001 said: I don't have such UI to toggle characters. I have a notepad. lol. The comfy generation face looks bad. It's also blurry. You can drag and drop forge image into comfy ui and see the forge workflow. tsukushi nova asian 1344x768 t2i. 1440p hiresfix upscale. Thank you for sharing your workflow. I am getting pretty close to generating it in Forge! Thank you. I remaked workflow based on your workflow. I changed K-Sampler because of Detailer. And Inpainting is independent (use Inpainting, change switch). This workflow need to add Lora Manager node to add Choth Loras easily. (and google translation node) To compare the upscaling methods A:[Simple Upscale] and B:[your HiresFix workflow], I can switch between groups A and B with a switch. Thanks to you, I was able to do T2I with ComfyUI for the first time! DOAXVV_T2I.json These pics made by this workflow. s2v.zip Edited February 24 by 5250 8
yhb-q Posted February 24 Posted February 24 On 2/19/2026 at 10:48 AM, 5250 said: Pos night,(1girl,solo,alone,solo focus,looking at viewer:1.4),embarrassed,cowboy shot,thighs,<lora:IL_DOAXVV:0.75>,__DOAXVV__,<lora:realistic filter [IL]:0.7>, lying on her stomach on a bed,ass focus,viewed from behind,pubic hair,, masterpiece, best quality, amazing quality, very aesthetic, high resolution, ultra-detailed,realistic, absurdres, scenery, photorealistic, depth of field, photorealistic details,realistic lighting,Cinematic composition,high contrast,moody and suspenseful ambiance,while the rest of the space fades into black, Neg anime, 3d, rendered, modern, old, oldest, cartoon, graphic, text, painting, crayon, graphite, abstract, glitch, deformed, mutated, disfigured, long body, lowres, bad anatomy, bad hands, missing fingers, extra digits, fewer digits, very displeasing, (worst quality, bad quality:1.2), bad anatomy, sketch, jpeg artifacts, signature, watermark, username, simple background, conjoined, bad ai-generated 👍 A10，A23，A25的下巴和肩膀的关系是错的吧 1
5250 Posted February 24 Posted February 24 11 hours ago, yhb-q said: 👍 A10，A23，A25的下巴和肩膀的关系是错的吧 Hello. I didn't notice it. There's definitely something wrong with her shoulder.
huehuehue9001 Posted February 24 Posted February 24 the face looks nice. doesn't look the same anymore. 1
huehuehue9001 Posted February 28 Posted February 28 (edited) Time for heresy. happy birthday. would you like some cake? Edited March 2 by huehuehue9001 7
huehuehue9001 Posted March 6 Posted March 6 (edited) cute misaki cosplay. Spoiler Spoiler looks weird without hands, redo. 3d is a bit hard, need to inpaint. the 2d just works. if you would like to upscale to 4k. going 1344x768 to 4k will not work well. 768p > 1440p (denoise 0.4) > 4k (denoise 0.15, cfg 2). if it doesn't work. you should give up or inpaint. Edited March 7 by huehuehue9001 5
x20man Posted March 6 Posted March 6 I would need to find better models, because none of them look like the ones in the game... 3
Recommended Posts
Create an account or sign in to comment
You need to be a member in order to leave a comment
Create an account
Sign up for a new account in our community. It's easy!
Register a new accountSign in
Already have an account? Sign in here.
Sign In Now