Name: Aesthetic Quality Modifiers - Masterpiece
Rating: 5 (11262 reviews)
Author: motimalu

We had an outage July 5th where users were not compensated, a fix has gone out.

July 5th Compensation

Krea 2 LoRA training is 50% off, and we're running a Krea 2 Training Contest with 1.8 million Buzz in prizes across four LoRA categories.

⚡ Krea 2 Training Contest + 50% Off Training

<h1 id="aesthetic-quality-modifiers-masterpiece">Aesthetic Quality Modifiers - Masterpiece</h1>Training data is a subset of all my manually rated datasets with the quality/aesthetic modifiers, including only the <code>masterpiece</code> tagged images.ℹ️ LoRA work best when applied to the base models on which they are trained. Please read the About This Version on the appropriate base models, trigger usage, and workflow/training information.Recommended prompt structure:Positive prompt (quality tags at the start of prompt):<pre><code>masterpiece, best quality, very aesthetic, {{tags}}, {{natural language}}</code></pre>

Trained on Krea 2 Raw base with diffusion-pipe. Strength of 1.5 is recommended when applied to the Turbo model.Previews generated with a slightly adjusted int8 quant workflow by Winnougan: <a target="_blank" rel="ugc" href="https://huggingface.co/Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8">https://huggingface.co/Winnougan/Krea-2-Base-Turbo-NVFP4-FP8-INT8</a>Workflow is in included in the preview images, you can drag and drop into ComfyUI.Same dataset as v5.1, adjusted captions to simply use "masterpiece, very aesthetic" trigger tags + NL captions for Krea 2 - which is assumed to have no booru tags knowledgeCan benefit from being used together with the krea2filterbypass LoRA: <a target="_blank" rel="ugc" href="https://civitai.red/models/2728234/krea2filterbypass?modelVersionId=3067151">https://civitai.red/models/2728234/krea2filterbypass?modelVersionId=3067151</a> See: <a target="_blank" rel="ugc" href="https://civitai.red/images/135097274">https://civitai.red/images/135097274</a>Training configs: <a target="_blank" rel="ugc" href="https://github.com/motimalu/diffusion-training-configs/tree/main/diffusion-pipe/krea/configs">https://github.com/motimalu/diffusion-training-configs/tree/main/diffusion-pipe/krea/configs</a>

v5.1 [krea2]

Trained on Anima Base 1Same dataset as v5.0 with a mix of natural language and tag captions.386 images, all masterpiece tagged images trained in Kirazuri (Anima) model version 2 dataset.Partitioned and trained at multi-res 1024, 1280, 1536Trained for 1,413 Steps, 3 Epochs. Training config:<pre><code># trained using diffusion-pipe commit b0aa4f1e03169f3280c8518d37570a448420f8be
# NCCL_P2P_DISABLE="1" NCCL_IB_DISABLE="1" NCCL_CUMEM_ENABLE="0" deepspeed --num_gpus=1 train.py --deepspeed --config anima-lora.toml --i_know_what_i_am_doing

output_dir = '/mnt/d/anima/training_output/anima-base-1-masterpiece-v51'

dataset = 'dataset-anima-masterpiece.toml'

# training settings
epochs = 3
# Per-resolution batch sizes
micro_batch_size_per_gpu = [[1024, 32], [1280, 24], [1536, 16]]
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1
warmup_steps = 30
lr_scheduler = 'cosine'

# misc settings
save_every_n_epochs = 1
activation_checkpointing = true
#reentrant_activation_checkpointing = true

partition_method = 'parameters'

save_dtype = 'bfloat16'
caching_batch_size = 1
map_num_proc = 8
steps_per_print = 1
compile = true

[model]
type = 'anima'
transformer_path = '/mnt/c/workspace/models/diffusion_models/anima-base-v1.0.safetensors'
vae_path = '/mnt/c/workspace/models/vae/qwen_image_vae.safetensors'
llm_path = '/mnt/c/workspace/models/text_encoders/qwen_3_06b_base.safetensors'
dtype = 'bfloat16'
#cache_text_embeddings = false
llm_adapter_lr = 0
#timestep_sample_method = 'uniform'
flux_shift = true
multiscale_loss_weight = 0.5
sigmoid_scale = 1.3

[adapter]
type = 'lora'
rank = 32
dtype = 'bfloat16'

[optimizer]
type = 'adamw_optimi'
lr = 4e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8</code></pre><pre><code>resolutions = [1024, 1280, 1536]

enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 9

# Totals
# 386 images
# 16 repeats from captions.json

# 153 images
[[directory]]
path = '/mnt/d/training_data/0_masterpieces_kirazuri/1536x1536'
resolutions = [1024, 1280, 1536]

# 44 images
[[directory]]
path = '/mnt/d/training_data/0_masterpieces_kirazuri/1280x1280'
resolutions = [1024, 1280]

# 189 images
[[directory]]
path = '/mnt/d/training_data/0_masterpieces_kirazuri/1024x1024'
resolutions = [1024]
</code></pre>

v5.1 [anima-base-1]

v5.0 [anima-preview-3]

v4.1 [anima-preview-2]

v4.0 [anima-preview]

Trained on <a target="_blank" rel="ugc" href="https://huggingface.co/Tongyi-MAI/Z-Image-Turbo">Z Image Turbo</a>With a much smaller dataset of 78 images to prevent degradation at higher step counts

v3.2 [z-image-turbo]

Trained on <a target="_blank" rel="ugc" href="https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B">FLUX.2-klein-base-4B</a>

v3.1 [klein-4b]

Trained on <a target="_blank" rel="ugc" href="https://civitai.com/models/795765/illustrious-xl">Illustrious v0.1</a>Included newly rated images.Dataset size: 332 images.Recommended prompt structure:Positive prompt (quality tags at the end of prompt):<pre><code>{{tags}}
masterpiece, best quality, very aesthetic</code></pre>Generation Settings:Previews are generated in Forge with upscaling and adetailer.For Noobai V-Pred, a ComfyUI workflow reference with DynamicThresholding, Upscaling, and FaceDetailer can be found here: <a target="_blank" rel="ugc" href="https://civitai.com/posts/11457095">https://civitai.com/posts/11457095</a>

v3.0 [illustrious]

Trained on <a target="_blank" rel="ugc" href="https://huggingface.co/Qwen/Qwen-Image">Qwen Image</a>Using <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe/blob/4a61bcfd82eb8d54a807ac7a15f548fab73ae5e1/examples/qwen_image_24gb_vram.toml">diffusion-pipe qwen config</a>Dataset resolution of 640Previews generated with the <a target="_blank" rel="ugc" href="https://huggingface.co/lightx2v/Qwen-Image-Lightning/blob/main/Qwen-Image-Lightning-4steps-V1.0.safetensors">lightx2v Lightning 4step LoRA</a>

v3.0 [qwen]

Trained on <a target="_blank" rel="ugc" href="https://civitai.com/models/833294/noobai-xl-nai-xl?modelVersionId=1190596">NoobAI-XL (NAI-XL) V-Pred 1.0-Version</a>Recommended prompt structure:Positive prompt (quality tags at the end of prompt):<pre><code>{{tags}}
masterpiece, best quality, very aesthetic</code></pre>Included many newly rated images and small training/tagging updates:<ul><li>change<code>min_snr_gamma</code> from 5 to 0<ul><li>should produce closer lighting/blacks to base model</li></ul></li><li>change <code>very awa</code> tag back to <code>very aesthetic</code><ul><li>reverted this change from previous version v1.0</li></ul></li><li>tag <code>nsfw</code> <code>explicit</code> with imgutils anime_rating</li></ul>Previews are generated in Forge with DynamicThresholding (CFG-Fix) Integrated enabled, settings:<pre><code>dynthres_enabled: True, dynthres_mimic_scale: 7, dynthres_threshold_percentile: 1, dynthres_mimic_mode: Half Cosine Down, dynthres_mimic_scale_min: 1, dynthres_cfg_mode: Half Cosine Down, dynthres_cfg_scale_min: 3, dynthres_sched_val: 1, dynthres_separate_feature_channels: enable, dynthres_scaling_startpoint: ZERO, dynthres_variability_measure: STD, dynthres_interpolate_phi: 1</code></pre>A ComfyUI workflow reference with DynamicThresholding, Upscaling, and FaceDetailer can be found here: <a target="_blank" rel="ugc" href="https://civitai.com/posts/11457095">https://civitai.com/posts/11457095</a>

v2.3 [noobai-v-pred-1]

[WAN 14B] LoRA (experimental)<ul><li>Trained with <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe">diffusion-pipe</a> on <a target="_blank" rel="ugc" href="https://huggingface.co/Wan-AI/Wan2.1-T2V-14B">Wan2.1-T2V-14B</a> with the same (image-only) dataset as v2.3 [noobai v-pred]<ul><li>Currently curating a video dataset</li></ul></li><li>Video previews generated with <a target="_blank" rel="ugc" href="https://comfyanonymous.github.io/ComfyUI_examples/wan/#text-to-video">ComfyUI_examples/wan/#text-to-video</a><ul><li>Loading the LoRA with LoraLoaderModelOnly node and using the fp8 14B: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/diffusion_models/wan2.1_t2v_14B_fp8_e4m3fn.safetensors">wan2.1_t2v_14B_fp8_e4m3fn.safetensors</a></li><li>Higher quality previews use the full fp16 14b: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/diffusion_models/wan2.1_t2v_14B_fp16.safetensors">wan2.1_t2v_14B_fp16.safetensors</a></li><li>Recommend following prompting guide for movement to avoid still images/jitter: <a target="_blank" rel="ugc" href="https://www.comfyonline.app/blog/wan2-1-prompt-guide">https://www.comfyonline.app/blog/wan2-1-prompt-guide</a></li></ul></li><li>Image previews generated with modified <a target="_blank" rel="ugc" href="https://comfyanonymous.github.io/ComfyUI_examples/wan/#text-to-video">ComfyUI_examples/wan/#text-to-video</a><ul><li>Setting the frame length to 1</li><li>Adding Upscaling</li></ul></li><li>Better results with text-to-image than text-to-video for this version (due to training on images only)</li></ul>

v2.0-alpha [wan-t2v-14b]

[WAN 1.3B] LoRA (experimental)<ul><li>Trained with <a target="_blank" rel="ugc" href="https://github.com/tdrussell/diffusion-pipe">diffusion-pipe</a> on <a target="_blank" rel="ugc" href="https://huggingface.co/Wan-AI/Wan2.1-T2V-14B">Wan2.1-T2V-14B</a> with the same (image-only) dataset as v2.0-alpha [WAN 14b]<ul><li>Currently curating a video dataset</li></ul></li><li>Video previews generated with <a target="_blank" rel="ugc" href="https://comfyanonymous.github.io/ComfyUI_examples/wan/#text-to-video">ComfyUI_examples/wan/#text-to-video</a><ul><li>Loading the LoRA with LoraLoaderModelOnly node</li><li>Recommend following prompting guide for movement to avoid still images/jitter: <a target="_blank" rel="ugc" href="https://www.comfyonline.app/blog/wan2-1-prompt-guide">https://www.comfyonline.app/blog/wan2-1-prompt-guide</a></li></ul></li></ul>

v2.0-alpha [wan-t2v-1.3b]

Trained on <a target="_blank" rel="ugc" href="https://civitai.com/models/833294/noobai-xl-nai-xl?modelVersionId=1190596">NoobAI-XL (NAI-XL) V-Pred 1.0-Version</a>Recommended prompt structure:Positive prompt (quality tags at the end of prompt):<pre><code>{{tags}}
masterpiece, best quality, very aesthetic</code></pre>With the kohya_ss dev branch, <code>v_paratemization</code> ,<code>zero_terminal_ssr</code> enabled, and noise offet set to 0.Included some newly rated images and small updates to match the noobai tagging:<ul><li><code>by {artist}</code> -&gt; <code>artist:{artist}</code></li><li><code>very aesthetic</code> -&gt; <code>very awa</code></li></ul>Previews are generated in Forge with DynamicThresholding (CFG-Fix) Integrated enabled, settings:<pre><code>dynthres_enabled: True, dynthres_mimic_scale: 7, dynthres_threshold_percentile: 1, dynthres_mimic_mode: Half Cosine Down, dynthres_mimic_scale_min: 1, dynthres_cfg_mode: Half Cosine Down, dynthres_cfg_scale_min: 3, dynthres_sched_val: 1, dynthres_separate_feature_channels: enable, dynthres_scaling_startpoint: ZERO, dynthres_variability_measure: STD, dynthres_interpolate_phi: 1</code></pre>

v1.0 [noobai-v-pred-1]

Trained on NoobAI-XL (NAI-XL) <a target="_blank" rel="ugc" href="https://civitai.com/models/833294?modelVersionId=1022833">Epsilon-pred 1.0-Version</a>Recommended prompt structure:Positive prompt (quality tags at the end of prompt):<pre><code>{{tags}}
masterpiece, best quality, very aesthetic</code></pre>