Big Love Klein3 is also offered for $10 on Tensor Art. You can get Big Love Eryn1, Ultra5, Photo6, Zuna1 or Zia1 for $5.50 and Gwen1 or Hyper1 for $10 on Fanvue. Click 'Follow', buy a post, check your Fanvue DMs! A subscription includes the latest release and new ones within 30 days. Any problem? Contact me via chat.
To create up to 85 free images per day in highest quality you can use Big Love here. For discussing photorealistic image generation join me on Discord.
Big Love Klein is a fine-tune of Flux 2 klein. It can do txt2img and also edit images with prompts. It only needs 4 steps, so is quite fast. Check the "About this version" box on the right for info on the individual variants. Some are up to 3x faster.
It is recommended to use it with ComfyUI. Put the downloaded model into the diffusion models sub folder of ComfyUI. You also need qwen_3_8b_fp8mixed.safetensors in the text_encoders sub folder (Qwen3-8B-Q5_K_M.gguf wth 8 GB VRAM) and flux2-vae.safetensors in the vae sub folder.
Settings are 4 steps (2 steps for upscaling), cfg1, Euler sampler, Beta scheduler, 832x1216, 1024x1536, 1280x1920 or 1536x2240 pixels
Big Love Gwen is a finetune of Qwen Edit 2511 and can do NSFW image editing. It produces more realistic and sharp images with better skin texture & genitals than other Qwen checkpoints/loras. Txt2img is less reliable, but generates nice results too.
Big Love Gwen1 allows commercial usage without additional fees (if no auto-generation involved) unlike Big Love Klein. With its better anatomy it is easier to get good editing results. As fast at 6 steps as Klein at 4 steps. Check the "About this version" box on the right for info on the individual variants.
It is recommended to use ComfyUI. The Gwen1 workflows offer a skin enhance feature that is not available in Forge Neo. Move the Gwen1 model file into the diffusion_model sub folder of the models folder of ComfyUI, qwen_2.5_vl_7b_fp8_scaled.safetensors into the text_encoder sub folder and qwen_image_vae.safetensors into the vae sub folder.
For training a lora read here for Gwen1.
Settings are 6 steps (3 steps for upscaling), cfg1, Euler sampler, Beta scheduler, 832x1216 or 1024x1536 pixels.
Big Love Eryn is a fine-tune of Ernie-Image(-Turbo). You can run it locally with ComfyUI and Forge Neo. The fp8 version works with 8 & 12 GB VRAM. Use the (pruned) bf16 version if you have 16 GB VRAM or more. The "Full" versions are only meant for training a lora (coming soon!). Put the downloaded model into the diffusion models sub folder of ComfyUI or the StableDiffusion sub folder of Forge Neo. You additionally need to place ministral-3-3b.safetensors in the text_encoder sub folder and flux2-vae.safetensors in the VAE sub folder.
Settings are 4-6 steps (3-4 steps for upscaling), cfg1, Euler sampler, Beta scheduler, 832x1216 or 1024x1536 pixels.
Big Love Hyper supports generating at image sizes of 1664x2232 (4 megapixel) and higher with SDXL. Can be easily upscaled to 9 megapixel unlike the normal low-res images. Hyper currently has not fully stable anatomy, so needs more runs. Some prompts work better than others. It produces highly detailed and life-like images close to real photos. It has a special quality that cannot be achieved by just upscaling. The Hyper versions of Big Love are pay-only and special license conditions apply (see below).
Settings are DMD2 lora with strength 1.0, LCM Exponential, 8 steps, cfg 1, clip skip 2, 1664x2432 or 1536x2240. Upscale with img2img with DMD2 lora with strength 1.0, same prompt, LCM Exponential, 1.25-1.5x, 4-8 steps, cfg 1, 0.3-0.5 denoise, clip skip 2.
Big Love ZT, Zia or Zuna are fine-tunes of Z-Image(-Turbo) on the Big Love dataset. You can run it locally with ComfyUI and Forge Neo. The fp8 version works faster with 8 GB VRAM but produces more noise so may sometimes look more realistic, but details are not as good. Use the (pruned) bf16 version if you have 12 GB VRAM or more. The "Full" version is only meant for training a lora. Put the downloaded model into the diffusion models sub folder of ComfyUI or the StableDiffusion sub folder of Forge Neo. To make Big Love ZT/Zia/Zuna work you additionally need to place qwen3_4b.safetensors in the text_encoder sub folder and ae.saftendors in the VAE sub folder.
Settings are 8 steps, cfg1, Euler or DPM++ 2s a RF sampler, Normal or Simple or Beta scheduler, 832x1216, 1024x1536 or 1280x1920 pixels. For upscaling 8 steps, cfg1, Euler sampler, Normal or Simple or Beta scheduler, 1.5x upscale.
Big Love Ultra supports generating at image sizes of 1280x1870, 1360x1984 and higher. Can be easily upscaled to 6 megapixel unlike the normal low-res images. More photorealistic and detailed. A special quality that cannot be achieved by just upscaling. Ultra produces more details & sharpness when upscaling and detailing low-res images. The Ultra versions of Big Love are pay-only and special license conditions apply. See below.
Settings are DMD2 lora with strength 1.0, LCM Exponential, 8 steps, cfg 1, clip skip 2, 1280x1864 or 1360x1984. Upscale with img2img with DMD2 lora with strength 1.0, same prompt, LCM Exponential, 1.25-1.5x, 4-8 steps, cfg 1, lora strength 1.0, 0.3-0.5 denoise, clip skip 2.
Big Love Photo and Insta1 are finetuned versions of XL. Lust1 is a finetuned version of Lustify merged with Photo. They were trained on thousands of images and dozens of new concepts. Why they are the most realistic and versatile. For more details and prompting Photo1 read this article.
Settings are DMD2 lora with strength 1.0, LCM Exponential, 8 steps, cfg 1, clip skip 2, 832x1216 or 1024x1496. Upscale with img2img with DMD2 lora, same prompt, LCM Exponential, 1.25-1.5, 4-8 steps, cfg 1, lora strength 1.0, 0.3-0.5 denoise, clip skip 2.
Big Love XL is a combination of 5 different training branches of SDXL: original SDXL, bigASP, NatVis, Anteros and Pony. With a bit of Pony in it, some Pony loras (poses, characters) work too. It outputs a more photorealistic & creative look than the Pony versions and can also do paintings & cartoon.
Settings are DMD2 lora with strength 1.0, LCM Exponential, 8 steps, cfg 1-1.5, clip skip 2, 832x1216 or 1024x1496. Upscale with img2img with DMD2 lora, same prompt, LCM Exponential, 1.5-2x, 8-12 steps (XL1/XL2), 4-8 steps (XL2.5-XL4), cfg 1, lora strength 1.0, 0.3-0.5 denoise, clip skip 2.
Big Love Pony produces different images than both Pony & SDXL-based models. Pony tags are understood, but interpreted a bit differently. Some images lean more in the one or other direction depending on the prompt. It works with Pony loras as well as normal SDXL ones. It can turn anime/cartoon into photorealistic images in img2img. Pony3 includes a bit of Illustrious.
Settings are DMD2 lora with strength 1.0, LCM Exponential, 8 steps, cfg 1-1.5, clip skip 2, 832x1216 or 1024x1496. Upscale with img2img with DMD2 lora, same prompt, LCM Exponential, 1.5-2x, 8-12 steps, cfg 1, lora strength 1.0, 0.3-0.5 denoise, clip skip 2.
ComfyUI Workflow:
Download one of the following images and drag'n'drop it onto Comfy:
Klein, Eryn and Gwen workflows are embedded in the showcase images.
Z-Turbo/Z-Base txt2img & Upscale Workflow (recommended)
SDLX DMD2 txt2img & Upscale Workflow (recommended)
SDXL DMD2 txt2ing & Upscale for Big Love Ultra
SDXL DMD2 img2img Workflow
SDXL Lightning Workflow
Trained Concepts
Newer versions of Big Love include the concepts of older versions, e.g. Photo4 can do everything that Photo1 to Photo3 can do and more. The only exception are the Insta concepts, but they have been partially added to never versions too.
Photo6/Ultra5: ai-style, bodyscape, butt plug, candid amateur, cumshot, illustrious-style, outdoor sex, partner fingering, photoart, innie pussy, ray-style, rebellious, reflection, sensual, sexy ass, sweetheart, wedgie
Photo5/Ultra4/Hyper1: alluring, anal sex, artistic photo, artistic pinup, charming, cum portrait, cute asian, cute pinup, cute portrait, double oral, dreamy-style, face pov, fantasy pinup, fantasy-style, kai-style, lowkey, mj-style, outdoor flashing, photo art, pole dance, public indecency, public nudity, rock climbing, scify-style, sword pose, vaginal sex
Photo4/Photo4.5/Ultra2/Ultra3: 2girls kissing, acrobatic sex, average face, beautiful face, amateur pinup, anal fingering, artistic portrait, cuddling, cum swapping, doggystyle pov, even skin color, fashion photo, female masturbation, female pov, finger licking, food porn, footjob, french kiss, full nelson, funny movie, funny tongue out, insta cute, legs pov, low contrast, mirror portrait, nude pose, oral pov, orgasm face, outdoor nude, penis licking, pinup photo, pro photo, pro portrait, pussy licking, sex pov, sexy portrait, vaginal fingering, waterfall
Photo3/Lust1/Ultra1: 360 degree photo, 3d selfie, 69 position, adorable, alt beauty, anal gape, anal sex, average face, beach life, beautiful, beautiful face, candidness, cum, cute, deepthroat, dildo insertion, double exposure, enthusiastic, extreme pose, faces, fashion photo, female pov, fisting, follow me pov, girl next door, gorgeous, handjob, heroin chic, huge feet, huge hands, huge nipples, light rays, light streaks, lighting, low/mid/high contrast, lowkey/midkey/highkey, low/mid/high saturation, middle finger, motion blur, object insertion, detailed eyes, teeth, pro photo, pussy gape, real life, rooftop, round ass, sad, sexy pose, skin texture, skin tone, spread pussy, squirting, street portrait, testicle licking, testicle sucking, vaginal sex, voluminous hair, water splash
Insta1: amateur photo, fashion photo, insta selfie, pinup photo, pro photo, real life, big ass, bimbo, cute, luxurious, adorable, enthusiastic, average face, beautiful face, detailed eyes, sexy legs.
Photo2: 3d selfie, 69 position, adorable, alt beauty , amateur photo, ball licking, ball sucking, beach life, bubble butt, low/mid/high contrast, dildo insertion, double exposure, enthusiastic, extreme pose, female pov, follow me pov, forced perspective, girl next door, gorgeous, handjob, heroin chic, huge feet, huge hands, huge nipples, light rays, light streaks, lighting, middle finger, motion blur, no tan lines, object insertion, pussy gape, rooftop, sad, low/mid/high saturation, sexy pose, squirting, street portrait, water splash.
Photo1: amateur photo, pro photo, lighting, lowkey, midkey, highkey, beautiful, cute, candid, real life, average face, remarkable face, voluminous hair, skin tone, skin texture, 360 degree photo, spread pussy, deepthroat, fisting, anal gape, anal sex, vaginal sex.
Training a (character) lora
My advice is this: Take 10-30 high quality images (at least 50-100 for styles or poses), download OneTrainer, click the SDXL lora preset, choose Big Love as the base checkpoint, add your images as a concept, add tags on the Tools tab, and start training. If you train on Big Love Ultra change the training resolution to 1536 and provide 1248x1832 images.
You will only get the full Big Love quality by training on real photos or extremely photorealistic images, which is usually not the case with a fictional character. It is all about image quality and not choosing boring images. Don't add posing images to a character lora, because Big Love provides the posing later for it. Focus mainly on good portraits with a few upper and full body images. Rather fewer higher quality images than a lot of average images. Read here for traing a lora on Big Love ZT2, ZT3, Zia1, Zuna1 or Eryn1.
Fixing Anatomy
If you want to keep the seed and repair anatomical problems in an image, activate the Extra checkbox in A1111/Forge and set Variation Strength to 0.01 to 0.1. Then generate until anatomy is fine. In Comfy there are sampler nodes that support variation seed (e.g. Inspire Pack), which do the same thing. Also possible to change steps or add commas to the prompt, but I do not recommend it. Extra/Variation Seed is more powerful and convenient as it gives you endless variations with just one click. There is always a variation which looks similar or even better with great anatomy.
You don't need negative prompts for this. They are rather ineffective. You can also change the positive prompt to suppress things. Anatomy problems are also created by the prompt if you prompt contradicting poses that the model cannot combine. Better to fix the prompt first then.
Reducing Saturation, Warm Colors and Sharpness
These SDXL loras reduces the saturation and change the color temperature without reducing image quality. The lora strength defines the effect. Negative values increase saturation or add more warmth. To make an image softer use a lower strength for the DMD2 lora, e.g. 0.95, 0.9, 0.85.
Big Love Hyper/Ultra/Photo Sizes
Big Love Hyper supports these image sizes:
1664x2432 (recommend)
1536x2240
Big Love Ultra supports these image sizes:
1360x1984
1280x1872 (recommend)
Big Love Photo supports these image sizes:
1024x1496 (recommend)
832x1216
You can also use the 16:9 or 2:3 versions of these image sizes as well as landscape orientation, but the above should be more reliable.
The higher the image size, the lower the probability of good results. If you get too many problems with a higher image size, switch to the lower one. Some prompts do higher sizes fine, some do not. Do not use a lower image size than the recommended one, otherwise image quality will be lower. With higher image sizes anatomy problems occur too often.
Remove fill words and unusual tags in the prompt. They can cause a quality reduction with Big Love Ultra/Hyper as they were not trained. Shorter prompts should work better as it is less likely that bad or untrained tags are in it.
SFW Output - No Tits Please
"Is there a trick so that Big Love doesn't always generate raised shirts or visible breasts?" I get this asked so often, while it is so obvious. Big Love knows what men want so it gives it to them. Negatives don't work that well for it. So you got to speak to it in the positive prompt like you were taking to a nun. "Breasts? How dare you speak such an obscene word to me." Don't mention breasts or any of these sexy parts. Instead describe clothes a lot. Don't write down your impure thoughts. God forbid! Yeah, I know, so hard. Describe her body shape, which also implies certain tits underneath the clothes.
XL Prompting Advice
SDXL prompts & prompts from various SDXL checkpoints work with it. Pony prompts too, but they create a more photorealistic look. Natural language prompts tend to create artistic glamour images with less skin detail while normal tags do more photo-style images. Big Love XL gives individual words more attention, so you don't need to rework prompts like Big Love Pony requires sometimes.
XL1/XL2 has a tendancy towards cuteness and bright skin. If you get too much of it, remove words like cute, sweet, pale/fair skin & flash from the prompt. It also creates realistical analog photos with reduced image quality. Sometimes better to remove terms like analog or grain to increase quality. Decrease cfg to get a more natural look.
XL2.5 can sometimes get too sharp. Reduce cfg or use less steps in img2img to make images softer.
Pony Prompting Advice
Pony prompts & prompts from other checkpoints work with it. Score tags and fill words like masterpiece or perfect skin have as good as no effect usually. A negative prompt is often unnecessary unless you want to make something vanish or avoid underage, but it can also change the style of the image.
Pure booru tags and some other samplers may produce Pony style. Photographic terms or a simple prompts ensure a photo look. Sometimes it refuses to produce medium shots, portraits or close-ups unless you weigh the term heavily or remove feet, shoes, high heels, boots etc. from the prompt. wavy hair sometimes does not look so great.
If it refuses to do porn, try simplifying the prompt. If you want to do a portrait, don't describe surroundings too much. If you want to do porn, describe the action mainly and don't focus on the people or surroundings in the prompt. The largest part of the prompt should be about the main subject of the image.
If a prompt does not produce the intended image, you often only need to modify it a bit to make it work perfectly.
Using Character Loras
Big Love Pony supports Pony characters, but they look like real humans. There were some complaints about them not working, but the problem was not the model itself so far. So here are some tips:
* Try Pony as well as SDXL character loras
* Use a lora weight up to 1.5 if it does not reduce quality
* If it is a well known character, add the name in the prompt and also give it a higher weight if necessary.
* Check your negative prompt for problems. Delete it if necessary and rewrite it from scratch.
* Render a lot of images to be able to pick those with the best resemblance.
More SDXL Photorealism:
You can highly improve the images generated with this model by using img2img, upscaling, detailing etc. The SDXL Lightning lora enhances the look and constrast despite only 8 steps. Alternatively, the SDXL DMD2 lora produces a sharp, but more natural look with less contrast. Also check out the Subtle Style loras, which nicely enhance images.
Big Love License
Big Love SDXL versions are licensed under CreativeML Open RAIL++M, the Big Love Z-Image, Ernie & Qwen versions are licensed under Apache 2.0, and the Big Love Klein versions are licensed under the FLUX Non-Commercial License with the following additional terms:
For clarity, the following version tiers are defined:
Restricted versions: Ultra, Hyper, Klein, Zia, Zuna, Eryn, Gwen
Open versions: Pony, XL, Photo, Insta, Lust, ZT
1. You may use Big Love without crediting the creator.
2. You may sell the images that it generates.
3. You may only run the official Big Love versions of the SubtleShader account on Civitai and Tensor Art. Running it on other public or shared servers commercially (including other Civitai and Tensor Art accounts) requires a separate license or permission from the creator.
4. You may not sell this model or merges of this model.
5. You may not merge, share merges of, or share LoRA extractions incorporating Restricted versions. You may merge and share merges of Open versions.
6. When sharing the allowed merges of Open versions, as well as finetunes of them, you must apply these same additional license terms.
7. Redistribution of Restricted versions is prohibited. They may only be obtained from official sources designated by the creator.
8. All use-based restrictions of the CreativeML Open RAIL++M license apply to SDXL versions and their derivatives. All use-based restrictions of the FLUX Non-Commercial License apply to Klein versions and their derivatives.
9. Any restrictions in these additional terms may be waived only with explicit written permission from the creator.
TLDR: This is basically the same license as before (just stated more clearly), so nothing changes for the Pony, XL, Photo, Insta, Lust and ZT versions of Big Love. However, restrictions were added for Ultra, Hyper, Klein, Zia, Zuna, Eryn and Gwen versions (no merging, no redistribution). The "Share merges" condition below the license on the right hand side was only deactivated because of the restricted versions. Merging is still allowed for the open versions.
Many thanks to the creators of SDXL, Z, Qwen, Flux klein, Pony, bigAsp and Lustify for making their wonderful models available. Big thanks to RaymondLuxuryYacht for allowing me to train on his fabulous images.
Description
Improved photorealism, better nipples & genitals than Big Love ZT1. Compared to Z-Turbo it has a more photorealistic look, more diverse faces, better skin (with the right prompt and seed), more variation between seeds and a Caucasian default look. You additionally need to place qwen3_4b.safetensors in the text_encoder sub folder and ae.saftendors in the and VAE sub folder. Works with ComfyUI and Forge Neo.
FAQ
Comments (59)
Is ZT2 more Loras friendly than V1?
You tell me... Will try training a lora later and compare to an existing one. Might be necessary to train directly on it. If so, will release a version that can be trained on with AI-Toolkit.
Just need to know if i can use my Ai model Lora before paying it..
@SubtleShader It will be great!
@realxgen If you send it to me I can try it.
Is it possible to run ZT2 with accelerators, like the DMD for SDXL, to make the image generation faster?
Actually ZT2 already has a accelerator similar to DMD2 already built-in, otherwise it would be 3x slower. There is a Nunchaku 4 bit version available of Z-Image-Turbo that probably runs 3x faster. I would have to convert ZT2 for Nunchaku to achieve the same.
Disadvantages: Nunchaku is probably similar in quality as fp8 maybe a bit less. So lower quality. It takes extensive cloud calculations with huge VRAM cards to convert and you need 2 Nunchaku versions für 5xxx as well as older gpus. Additionally, Forge Neo currently does not support Nunchaku for Z-Image-Turbo, only for SDXL. So it would be necessary to use ComfyUI.
I will look into it...
I tried TwinFlow Z Image, with 2 and 4 steps, its quality is terrible even with a custom node made for it
@qek I don't know TwinFlow but 2 or 4 steps is not enough for Z Turbo.
@SubtleShader TwinFlow was intended to use with 2 steps and more, I dislike it, even with a proper(?) sampler (node) made for TwinFlow Z. Indeed 2-4 steps isn't enough for Turbo, I just mean TwinFlow made for Z Image Turbo, check it
@qek I never use a speed lora or turbo checkpoint with anything below 8 steps. Don't need to check it to know that less does not work well.
@SubtleShader Yes, TwinFlow and the unoffical nodes for it are trash, forget it, I just tried to let the OP (H4rryMass) know it's possible, but Z Image is already accelerated (Turbo) and it's enough, no need to use something worse and cheaper. Again, forget it, I said that for the OP
Can someone teach me how to edit the area outside the face? I want to keep the face.
Its an SDXL model. So use Fooocus and inpaint. Fooocus has best inpainting.
Okay, I'm trying not to like Z-Turbo, but you are making it very difficult my friend! I DL'ed the latest ZT2 and I really like it... ARGH! Now I have to create all new Loras!! Great work, as always!! I posted a few images, just getting started, but I'm impressed for sure!
Thanks for posting the images! I like them. Training takes 2.5x to 3.5x longer. Z will only get better. Z-Turbo is less effective to train. Z base will become even better.
@SubtleShader Cool! I'm glad you're on our team! :-D
@GenoMachino Thanks! I plan to train my whole Big Love data set on Z-Turbo.
what does it take for you to 'like' a model type? gooning material? might wanna branch out...
@Narz Ooo... burn...
Big Love ZT2 Character Lora Test
What works best:
1. Train on ZT2, use with ZT2 (75% hit-rate) 👍
2. Train on Z-Turbo, use with Z-Turbo (50% hit-rate)
3. Train on ZT2, use with Z-Turbo (faded & grainy)
4. Train on Z-Turbo, use with ZT2 (very bad) 👎
Download the fp32 diffusers version of Big Love ZT2 above to train a lora on Big Love ZT2 with AI-Toolkit. Unzip the 16 GB zip file (30 GB unpacked) and point AI-Toolkit to the unzipped folder instead of Tongyi-MAI/Z-Image-Turbo. Train with training adapter, not de-distilled.
How many and what size images did you train with? Did you use a consistent resolution or a variety of resolutions and just let the trainer do its thing?
Is this the same version of ZT2 that I've downloaded from Tensor? I'm a little confused as the link above goes to the main CivitAI models page...
@boomboom1 24 images. All were above 1024 pixels so I trained 1024 resolution. I always use a total variety of image sizes.
@GenoMachino I only uploaded the bf16 version to TA. Yes, the link above is the fp32 download at the top of this page. It is the same zip file that I sent you.
@SubtleShader Yeah, sorry, I'm just not getting to a download page... that link keeps resolving to the main CivitAI.com page.. I can see the details in the link address, but just not able to get to a downloadable zip file...
@GenoMachino Sent again. Sorry.
How do you point Ai toolkit to a model, i have been using the auto downloaded model so far
@Ceylon_Ai Just paste or write the file path to it, e.g. C:\User\Ceylon\Download\BigLove_ZT2_Diffusers\, in the box where Tongyi-MAI/Z-Image-Turbo is written.
@SubtleShader Can that model we point to also be in .safesensor format? Or does it have to be in "diffuser file" format?
@GenoMachino AI-Toolkit currently only supports Z diffusers format. Otherwise I would not need to upload it here.
@SubtleShader Ha! Oh yeah, good point!! :-D Thanks!!
While other model authors couldn't care less about the convenience of LORA training, you not only considered this aspect but also wrote such a detailed LORA training manual. You're truly kind and responsible!
Hello,
I tried training, but I’m getting random results with the checkpoint. I’m probably doing something wrong. (29 photos to the dataset)
Do you select only 1024 during training, or also 512 and 768?
What learning rate and number of steps do you recommend?
Thank you for sharing.
@adridu51120 For higher quality only train 1024px. Do 100 epochs, save every 10th epoch and compare the last few ones. 0.0001 learning rate is fine. 29 images * 100 epochs / batch size 1 = 2900 steps. Less steps with higher batch size. For higher batch sizes multiply the learning rate with sqrt(batch size).
@SubtleShader
Thank you for the reply.
Do you enable EMA or DPO?
For captioning, is it better to use just the token, or to add more detailed descriptions?
Did you leave the other settings at their default values in AI-Toolkit (linear rank, caption dropout, etc.)?
I’m going to run some tests. :)
@adridu51120 No EMA & DPO. Tag each image. Keep other settings unless you have a reason not to.
@adridu51120 Good "how to" video on AI-Toolkit... https://www.youtube.com/watch?v=zegMOOKy_6A
@GenoMachino
So I know how to use AI-Toolkit, but what I’m looking for are the best settings to make a ZIT NSFW LoRA. That’s why I was asking the previous questions, to get the best settings with BigLove ZT2. :)
@adridu51120 for Z turbo training (AI Toolkit):
Adafactor, rank 32 or 16, batch size of 2 OR if your setup can't handle it use gradient accumulation 2 - it does the same job but slower and requires less vram, lr: 0.000141421356237, weighted, no quantization for the transformer, 8-bit for the text encoder, no captioning (only triggerword).
dataset:
~25-50 HIGH QUALITY images. 512 resolution is enough for Z image, I've trained with 768, 1024 and mix of them and they didn't give any better results for me.
1. Crucial for the best likeness: 1:1 headshots from multiple angles cropped from high resolution images (ideally not actual close-ups since the camera lenses usually distort the facial features, consistency is the key in these, if you can use images from same photoshoot).
5-15 of these.
2. Rest of the dataset should be a mixture of knee-up, chest-up pictures and just a few full body shots.
Prepare and crop the aspect ratio bucket groups with a tool. Each bucket should contain paired amount of images.
Steps: dataset size * 100 to 150 / 2.
Use the lora with 0.75 strength.
By the way has anyone succesfully used the biglove zt2 diffusers in Onetrainer? If so, let me know what setting to tweak to get it going. I tried to change the Base Model setting to mode_index.json in the biglove zt2 diffusers folder but I get an error message: NotImplementedError: Loading of single file Z-Image models not supported. Use the diffusers model instead. Optionally, transformer-only safetensor files can be loaded by overriding the transformer.
@virustotal Thanks for your recommendations! Turbo with adapter training or base training too?
@SubtleShader I haven't tested the base yet but I doubt you need to change anything, just stay at 100 epochs.
@virustotal Any tips for z image base training
@virustotal Hello, first of all thank you for sharing your training recommendations. I followed your advice and cropped my photos to a 1:1 ratio. However, I can’t set the batch size to 2 or the gradient accumulation to 2 because my VRAM maxes out, even with low VRAM enabled. I’m using an RTX 4080 Super.
I’ve just started a new training run with this dataset using the default settings to see how it turns out:
z-image turbo with turbo trainer adapter v2, LR: 0.0001, Adam8bit, float8 quantization.
I’m a bit mixed about the test results. I feel like I get better results when I test my LoRA with a different checkpoint than biglove and with a different VAE (ultra_flux).
By the way, Big Love ZT3 with base training coming soon... ZT3 will have a better quality too.
@adridu51120 in AI toolkit enable Low Vram and Layer Offloading settings at the top of the page and set both sliders to ~50%. It offloads half of the model to CPU RAM. I have 16 GB VRAM as well so it should be doable. During training you can check the VRAM usage and adjust the sliders to optimize the VRAM usage. If you don't need to have the optimal settings you can use the 8bit quants and 8bit optimizer, they can still produce decent results. The dataset is the most important part.
For Z base training I recommend using OneTrainer, BF16, prodigy_adv optimizer, learning rate 1, CONSTANT, LOGIT_NORMAL, and 100 epochs. In fact I would recommend it for Turbo as well but AI toolkit doesn't come with prodigy_adv and I haven't bothered to figure out how to use the ostris adapter in Onetrainer.
@SubtleShader YAY!! Can't wait!
@virustotal Hello, so you are doing training without a Training Adapter?
Generation only... lame, tbh. I'm sure you paid for the rights to use all the images you used in training, right?
nah he mixed garbage sdxl images in the training for sure to increase the dataset, it's just plain bad
It is pretty funny to me when people try to apply copy protection and the like to something stemming from what was basically the single biggest act of copyright and IP violations in the entire history of the universe.
@hcommand68990 Yes exactly, so you shouldn't have a hard time understanding why it's stupid that these people are being greedy with their shitty models trained on said material.
guys! I need your help. I may not understand something, but I can't find a normal workflow for the Photo 4.5 model. All the images are getting kind of vague. After reading a lot of information, including here on the creator's page, I still didn't understand which workflow works best. Some users have excellent quality images in the gallery, but unfortunately there is no metadata in them. Because of this, it is difficult to understand how they achieve such results. I use Comfyui on my home PC.
Use the workflow embedded in this image: https://civitai.com/images/49514234
It is also linked in the description.
Most images have metadata. But it is mainly Forge metadata and not ComfyUI one. You just need to copy and paste the prompt to reproduce. The seed is different between Comfy and Forge.
Also install Forge or Forge Neo. You won't regret it. Less fuzz and faster results.
waste of 10k buzz,
base z image turbo makes better images with lenovo ultrareal
compared to this model with lenovo
lenovo is for the base model. And why anyone needs that crap look? Its good to cover, if a model is not realistic, but otherwise?
if I may add, base model still makes better images than this finetune in many scenarios,
no use shilling for the guy, my guy
@MaxPJ Im not shilling:) I know the base is better, because lora merging, concept merging doesnt work with a distilled model, like with a normal base model. Yes, so its no use finetuning a distilled model.
What i said, why do you need Lenovo at all? You can prompt that look easily without crappify your results.