WakiStorage01

You can see my works' full generation data here -> https://www.seaart.ai/user/ab06e18b8ffee1f3e6fb917209a39b30 I create all my works here -> Seaart.ai. I no longer create in SD1.5 and PonyXL mediums, so I don't know how relevant my guides for them are by now. * If you feel like thanking me for one or more guides I produced. Or maybe a particular artwork. ( ͡~ ͜ʖ ͡°) * -> https://www.patreon.com/WakiStorage01 Mini updates section. 11.13.2024 - New checkpoint to use -> Anime Model OtakuReliableEnable (AMORE). It handles anime style better and seems to deep-fry output images less frequently too. 1.27.2025 GUIDES UPDATE. - added IllustriousXL guide. - added PonyXL characters pairing complementary guide. - Discovered that there are only so many characters can be put into profile description here on Pixiv and my museum of older guide updates now forever gone. 😭😭😭 ( and a little bit of now archived SD1.5 guide too ) -------------------------------------------------------------------------------------------------- *Turn on the "Legacy Mode" first*. I create all my works with it active because I'm very used to the UI. ILLUSTRIOUSXL GUIDE. ControlNet -> SDXL Model: WAI-NSFW-illustrious--SDXL LoRA: any IllustriousXL of your choosing. Weight is usually between 0.9-1 Upload an image for the reference Basic settings: don't touch anything except desired quantity of images. Other parameters will automatically adjust themselves after you upload the image. ControlNet: Type -> Canny Preprocessor -> Yes Preprocessor -> canny Preprocessing Model -> diffusers_xl_canny_full Preprocessor Resolution -> 1024 Control Mode -> Prioritize Prompts Control Weight -> 0.6-0.9 Advanced Config: Negative prompts -> doesn't seem to need any. I add character's default clothes there if they keep popping up when I change character's attire. I also add: 3D and realistic / semi-realistic in here and undesirable camera angles, mostly "from behind". Embedding -> I used to add "badhandv4" here but it seems to got deleted or something, leave it empty. VAE -> sdxl_vae. NOTE: I saw people recommend using none. I personally still use sdxl_vae. LCM -> off Sampling Method -> DPM++ 2M Sampling Steps -> 30 CFG Scale -> 5 - 7 NOTE: I saw it's recommended to use 5 almost unanimously. I use 7 still. Seed -> Random, -1 Clip Skip -> 2. NOTE: Again I saw some recommend using 1. I use 2 Positive prompts -> LoRA's trigger word + clothes, actions, environment, etc of your choosing. I use "anime screencap" when output image doesn't look anime enough to me and these three: "masterpiece, best quality, amazing quality". In conclusion: IllustriousXL is very similar to work with to PonyXL, except you use IllustriousXL LoRAs and Checkpoints and don't need to add the "score_9", "score_8_up", "score_7_up", " source_anime" string of prompts. IllustriousXL pros: better quality, particularly of small details and lighting. Much better style preservation. Seems to give less weird anatomy duds as well. IllustriousXL cons: Most LoRAs are very stiff, hence, higher canny control Weight needed. Some of its LoRAs are very sensitive to angle prompts, i.e. "from side", "from below", etc and tend to overdo it. "Where are the FAQ and advice and recommendations sections for IllustriousXL?" Answer: Except for some LoRAs being incredibly stiff, I didn't really encounter any notable troubles, so far. ------------------------------------------------------------------------------------------------------------------------------------------ PONYXL GUIDE. ( includes characters pairing complementary guide at the end of it, right before now archived SD1.5 guide ) ControlNet -> Canny -> upload a photo for the reference Mode -> SDXL Model ( Checkpoint ) -> AutismMix SDXL or Anime Model OtakuReliableEnable (AMORE) LoRA -> of your choice. LoRA's weight: 1 ( almost always 1, check the LoRA's Civitai page to be sure. ) Refiner -> Off Basic setting: don't touch anything except image quantity depending on how many you want AI to generate. Default is 1. ControlNet: Type: Canny Preprocessor: Yes Preprocessor: Canny Preprocessing model: diffusers_xl_canny_full Preprocessor resolution: 1024 Control mode: Prioritize Prompts Control Weight 0.25-0.45 ( I use 0.3 most of the time ) Advanced config Negative prompts: You can keep it pretty much empty. I add there only "blurry, monochrome, from behind, from above, from side, from below" for first work and "blurry, monochrome, from behind, from above" for the second one + characters' default clothing pieces when changing their outfits. VAE: sdxl_vae LCM -> off Sampling method: DPM++ 2M Sampling steps: 31 SFG Scale: 7 Seed: -1 Clip skip: 2 Positive prompts: score_9, score_8_up, score_7_up, source_anime, anime screencap, + *CHARACTER'S TRIGGER WORD and CLOTHES* + *ACTION AND ANGLE* Generate. --------------------------------------------------------+ Img2img section: You can salvage deep-fried ( doodle ) images by running it through img2img. Here's how: 1. Download the deep-fried image. 2. Replace the reference image with it. 3. Click img2img ( bottom right ) 4. Set denoising strength to 0.8. You don't need to change anything else. 5. Generate. Be advised, AI still can and will deep-fry the image from time to time here too. The solution to that is to add or subtract one sample step and attempt to generate again. And yes, despite how irrecoverable an image can look color-wise, AI is able to save it. Check out my recent posts on SeaArt, I now use img2img a lot and many reference images in my works are deep-fried ones. Click "Remix" and see for yourself. You can also use the img2img instrument to: 1. Fix small details that ruin otherwise good image. - let's say you got an almost perfect image but there's one or two small details that ruin it. Same as previous, download the image and run it through img2img with denoising strength 0.8. - additionaly, you can apply manual fixes to that images first before using img2img to further increase the AI's output accuracy. 2. To attempt to make an already good image even better. This is optional and changes are often very trivial. - Just download that image and replace your reference image with it, click img2img, set denoising strength to 0.8 and generate. One last important thing about img2img instrument. You can't use it like Canny. The reference image used there must closely resemble the character you're trying to create an image of in order for it to work. --------------------------------------------------------- Advice and explanation section: Always put this chain of prompts first in the positive prompts section before anything else -> score_9, score_8_up, score_7_up, source_anime <-. This is PonyXL's equivalent of "masterpiece, higres, very high quality, etc." prompts, but much shorter and standardized for every generation. --- PonyXL FAQs section: "AI turned my image into a doodle ( like this https://www.pixiv.net/en/artworks/121358783 ), what to do?" - There's no fixing that as far as I can currently tell. AI just does it once in a while. The next best thing you can do about it, since fixing Is not an option, is to copy the doodle's seed ( circular icon with "i" in the center is where you find the seed number ) and put into the seed section instead of -1 when trying to generate again. I recommend you doing this only if the doodle you've got looks like it would've otherwise been a decent image. If it does not, just generate again with -1. You can get multiple doodles in a row. I don't know why AI does it. Here's how to significantly reduce the chance of AI turning your next generation into a doodle after you got one. Repeat the generation attempt but this time add or subtract 1 sampling step. Yes it sucks to manually tweak sampling steps each time you get a doodle, but I simply haven't found a better solution yet. It's also not guaranteed to work, mind you, the success rate is ~4/5 times. "AI deep-fried my image or turned it into a static / rorschach painting. Why?" Two possible cases: 1. You used sampling method other than euler a. Switch your sampling method to euler a. I had this problem at the beginning when I was just starting learning PonyXL. Once I switched to euler a, the problem ( so far ) never appeared again. Also use this if DPM++ 2M turns every image into a static, not a doodle, the static. 2. Prompt strength modifiers. Some LoRAs react badly to attempts to both increase or decrease the strength of a prompt, I.e. (PROMPT:1.1) or (PROMPT:0.9). If AI started turning your images into a static after you added either, it's 100% the case why. PonyXL characters pairing guide. 1. Create an image of each character separately. 2. Stitch these two images manually, do some minor fixes/adjustments if necessary. 3. Copy the prompts from one of the character's creation message from SeaArt Bot, except these: "score_9, score_8_up, score_7_up, source_anime, anime screencap, 1girl, solo" 4. Go to the creation ControlNet page of the other. ( click small circular icon to the right ) 5. Add the LoRA of the character you copied the prompts of earlier. Keep both characters' LoRA strength the same. 6. Put the image with both characters you made early into the reference image window of the ControlNet. 7. Change "1girl" to "2girls", get rid of the "solo" prompt, add the "BREAK" prompt after "2girls". Add another BREAK prompt after the very last prompt. 8. Add the prompts you copied from the other character after the BREAK prompt you just added. 9. Click on send to img2img. 10. Set the Denoising Strength at 0.6, don't change anything else after that. 11. Generate. *You can set the Denoising Strength as low as 0.4 if characters' features bleed into each other is too severe. When all said and done your prompts window should look like this: "score_9, score_8_up, score_7_up, source_anime, anime screencap, 2girls, BREAK, UbelSSnF, green hair, side ponytail, purple eyes, black dress, halterneck, sleeveless, short dress, armlet, arm strap, badhandv4, arm behind head, armpit, black background, closed mouth, looking at viewer, smile, smug, licking lips, BREAK, LaufenSSnF, orange hair, short hair, hair bun, brown eyes, tassel, china dress, purple dress, sleeveless, halterneck, badhandv4, arm behind head, armpit, black background, closed mouth, looking at viewer, smile, blush, half-closed eyes" It would probably work for 3 different characters as well but I didn't test that and thus, can't attest. The only 3 characters in one image work I did so far was Helltaker's Cerberus and there I only needed a stitched image, one LoRA, 3girls prompt and no BREAK prompts. What's the "BREAK" prompt? From my understanding it helps AI to see where it needs to separate stuff more clearly, thus minimizing LoRAs or completely different things within image bleeding into each other. Multiple more advanced explanations of it exist on the Internet about how it organizes prompts into substacks for better image output accuracy. Easy to find if you want to. ------------------------------------------------------------------------- SD1.5 GUIDE. [ THE MEDIUM IS NO LONGER BEING USED BY ME. I DON'T GUARANTEE THE GUIDE RELEVANCY ] --------------------------------------------------------------------- The models ( Checkpoints ) I use: Step 1 - Sweet-mix v2.2-flat. Step 2 - AAM - AnyLoRA Anime Mix - Anime Screencap Style Model - v1 ------------------------------------------------------------------------- Here's how I do it and what settings I use: --- STEP 1 --- ControlNet -> Canny -> upload a photo for the reference Mode -> Default Model ( Checkpoint) -> Sweet-mix v2.2-flat. LoRA -> of your choosing. Basic settings: Image quantity -> 1-4 Image mode -> standard Image size -> should be already there and have the same dimension as your reference photo but smaller in scale. You can manually change it, but I don't recommend doing so. You'd get either monstrous anatomy or the character would be partially out of frame. ControlNet: Type -> Canny Preprocessor -> Yes. Control mode - > Prioritize pre-processing. Preprocessing model -> There's only one to choose. Preprocessor resolution -> 768 Control mode -> Prioritize pre-processing image ControlNet weight 0.4 - 0.45 ( you can go all the way up to 0.65, it's an option, but not every LoRA / checkpoint combination would allow that without the penalty of merging with the reference image. The higher the weight the higher the chance of the reference image bleeding into the generated image. Advanced config: Negative prompts -> check the negative prompts section. Embedding -> leave it empty VAE -> kl-f8-anime2 LCM -> off Sampling method -> DPM++ 2M Sampling steps -> 20 CFG Scale -> 7 Seed -> -1 Clip skip -> 2 PROMPTS SECTION. LoRA prompts - always go to its civitai page and copy them from there, as well as see the LoRA's recommended weight. It's usually 0.8 - 1. Pose/Angle prompts: This you'll have to write yourself depending on what you want the character of your choosing to do/show. Here's what I use: Both arms up - negative prompts: EasyNegative, (backlighting:1.4), washed away colors, dull colors, drawn by bad-artist, sketch by bad-artist-anime, multiple girls, lowres, text, error, missing arms, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, out of frame, blur, blurry, 3D, realistic, (open mouth:1.2), from above, from side, from below. Both arms up - positive prompts: 1girl, solo, anime screencap, frontlighting, (simple background, black background, dark background:1.3), masterpiece, absurdres, hdr, soft light, best quality, detailed, highres, shiny skin, shiny hair, (looking at viewer, eye contact with viewer:1.5), smile, smug, (closed mouth:1.2), arms up, raised arms, armpits, in the center, symmetrical, upright, One arm up - negative prompts: EasyNegative, (backlighting:1.4), washed away colors, dull colors, drawn by bad-artist, sketch by bad-artist-anime, multiple girls, lowres, text, error, missing arms, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, out of frame, blur, blurry, 3D, realistic, (open mouth:1.2), from above. One arm up - positive prompts: 1girl, solo, anime screencap, frontlighting, (simple background, black background, dark background:1.3), masterpiece, absurdres, hdr, soft light, best quality, detailed, highres, shiny skin, shiny hair, (looking at viewer, eye contact with viewer:1.5), smile, smug, (closed mouth:1.2), arm up, raised arm, armpit, (from side from below:1.5) Start generating. You'll be getting low-res raw images. Generate until you get a low-res image that resembles what you want the most. Download that image. --- STEP 2 A --- Model ( Checkpoint ) - AAM - AnyLoRA Anime Mix - Anime Screencap Style Model - v1 LoRA - the same you used for the step one. Use ControlNet - "Tile Resample". Upload the image you got from the step 1. Preprocessor - Yes. Control mode - Prioritize preprocessor. ControlNet weight 1. Image mode - Ultra High Quality. Beyond that, everything else stays the same, you don't need to change anything. --- STEP 2 B --- ( NOTE: I haven't used it in months, I don't know how relevant this still is. ) For the sake of speed, simplicity and comparison, do the --- STEP 2 A --- first. After that click repeat then send to img2img. Denoising Strength 0.5-0.65. You don't need to change anything else and your low-res raw image from the step 1 should be already there. Generate. You'll get one of these three outcomes: 1. Both A and B outputs are good, but style is different, either slightly or considerably. 2. Both A and B outputs are good and there is no difference between the two, or too small to notice. 3. One output is good and the other is bad. Just generate until you get what you want. Be patient, it's a numbers game. You might get what you want on your very first attempt or you might have to generate 10+ times before you finally get a good image. ------------------------------------------------------------------------- EXPLANATION AND ADVICE SECTION. I changed the model I use for Step 1 to Sweet-mix 2.2-flat because it gives way less bad anatomy and then, during Step 2, the AAM - AnyLoRA Anime Mix v1 model synergizes with it very well. The output is very consistent style/quality-wise. What are those "rare exceptions" I mentioned at the beginning? Sometimes LoRA is made in a way that it produces good images only when used with the model ( checkpoint) it was trained on. It's exceptionally rare from my experience. If you encounter situation like this, simply use the model it was trained on. LoRAs' creators almost always mention the model they used. You can mix and match models ( checkpoints ) for Step 1 and 2 to get unique end results. (word/phrase:1.5) <- Strength adjustment with which AI pays attention to a particular prompt. You can do it with multiple words/phrases at once, like this - (simple background, black background, dark background:1.3). It works both for positive and negative prompts. 1 is default, 1.5 is the safe upper limit adjustment at which AI is *PRACTICALLY* guaranteed to draw what you want ( 9 out of 10 times ), without doing weird things. Adjusting any higher might cause it turning image into something Vincent van Gogh would be proud of. Adding "anime coloring" prompt is usually the last resort option and most of the times you're better off trying different LoRA. Unlike "anime coloring", the "ANIME SCREENCAP" prompt seems to help out a lot with generating anime style images consistently. ---------------------------------------------------------------------------------------------------------- FAQ SECTION. "I got a good STEP 1 image, but STEP 2, both A and B turns it into a bad one, what to do?" - Change Image mode from "Ultra High Quality" to "High Quality". "During the step two, the image I've got has too many details / overly bad anatomy." - Drop the LoRA's strength by 0.1 and try to generate again. Keep repeating until the problem goes away. "During the step two, the image I've got is too sharp / the colors are oversaturated." - Drop CFG scale by 0.5. Keep repeating until the problem goes away. "The character I keep getting looks nothing like on Civitai's pictures / in the anime." - try to use the model ( checkpoint ) Lora's creator used to train it on. If the Lora's creator didn't specify it in the Lora's description, click example images on Civitai and see it there. ( NOTE: sometimes doesn't show, mostly when the image was "generated through an external generator". ) "The character wouldn't strike the pose I want." - Drop the LoRA weight by 0.05, increase the control weight by 0.05 and try again. Keep repeating until the problem is fixed. "The character's clothes wouldn't change to the ones I want or only partially." - Decrease the LoRA's weight by 0.1, put the character's default clothes prompts into the negative prompts section and enclose them within parentheses with up to 1.5 strength modifier. Example - (jacket, white shirt, necktie, formal clothes:1.5). Do the opposite for the clothes you want the character to wear. "Why "CHEEKY" fix?" - Because it doesn't work consistently. 3 things might happen: 1. It worked and you get the desired image. 2. It didn't work and you got the similar image with the same flaws or very similar to it. 3. It didn't work and gave you completely different image. You also have to remember to change seed back to -1 before the step two. More on the cheeky seed fix below in the "Cheeky fix and manual fixing section" "Why do your works differ by style and quality?" First - I haven't yet learned how to make them all uniform and don't know if it's even possible with the instruments I use. Second - Not all LoRAs are created equal. Some are very flexible and high quality - absolute joy to work with. Others, not so much. Often, LoRAs carry with them a particular drawing style that is impossible to fully get rid of. If it's possible, I don't know how, yet. ( NOTE: I finally figured out how to keep works more or less consistent quality/style-wise ) Same goes for pose and angle. Not all LoRAs are flexible enough for AI to draw them the way I want. It's pointless to attempt to brute-force AI by increasing the ControlNet weight because you'd simply get a monstrous anatomy. ---------------------------------------------------------------------------------------------------+ Cheeky seed fix and manual fixing SECTION. ( NOTE: Haven't used cheeky seed fix in months, I don't know how relevant this still is. ) Cheeky seed fix - use this when you got a great step one image, but it has just one or two details badly off that ruin it. Click on this image, there's the "seed section" in it. Copy the string of numbers there. Repeat the generation process, but this time add that string of numbers instead of -1 in the seed section. IMPORTANT - DO NOT CHANGE ANYTHING ELSE EXCEPT FOR THE SEED, OTHERWISE IT'S GUARANTEED NOT TO WORK. Also, don't forget to change the seed back to -1 again before the step two. It caught me off guard so many times at first and will catch you too. As mentioned in the FAQ section above, it might not work. I don't know why sometimes it does, sometimes it doesn't. The cheeky seed fix does not work for the step two. If t