Jalankan secara tempatan
spo_sdxl_10ep_4k-data_lora_webui.safetensors
355 MB · SafeTensor
Muat turun disediakan oleh Civitai. Sesetengah pencipta memerlukan log masuk Civitai dahulu.
Tentang model ini
STABLE DIFFUSION XL 1.0 · LoRA
SPO aka Step-Aware Preference Optimization by Rockeycoss is similar to DPO, the goal being to align models to human preferences and not just accuracy. As such, this lora acts as a powerful positive prompt. Add a positive weight to control the effect. The sweet spot will depend on how many other loras you have in play and what base model you are using, no additional information was given from the creator
Nota pencipta
Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference
Abstract
Generating visually appealing images is fundamental to modern text-to-image generation models. A potential solution to better aesthetics is direct preference optimization (DPO), which has been applied to diffusion models to improve general image quality including prompt alignment and aesthetics. Popular DPO methods propagate preference labels from clean image pairs to all the intermediate steps along the two generation trajectories. However, preference labels provided in existing datasets are blended with layout and aesthetic opinions, which would disagree with aesthetic preference. Even if aesthetic labels were provided (at substantial cost), it would be hard for the two-trajectory methods to capture nuanced visual differences at different steps.
To improve aesthetics economically, this paper uses existing generic preference data and introduces step-by-step preference optimization (SPO) that discards the propagation strategy and allows fine-grained image details to be assessed. Specifically, at each denoising step, we 1) sample a pool of candidates by denoising from a shared noise latent, 2) use a step-aware preference model to find a suitable win-lose pair to supervise the diffusion model, and 3) randomly select one from the pool to initialize the next denoising step. This strategy ensures that diffusion models focus on the subtle, fine-grained visual differences instead of layout aspect. We find that aesthetic can be significantly enhanced by accumulating these improved minor differences.
When fine-tuning Stable Diffusion v1.5 and SDXL, SPO yields significant improvements in aesthetics compared with existing DPO methods while not sacrificing image-text alignment compared with vanilla models. Moreover, SPO converges much faster than DPO methods due to the step-by-step alignment of fine-grained visual details. Code and model:
Model Description
This model is fine-tuned from . It has been trained on 4,000 prompts for 10 epochs. This checkpoint is a LoRA checkpoint. For more information, please visit
Citation
If you find our work useful, please consider giving us a star and citing our work.
{liang2024step,
title={Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization},
author={Liang, Zhanhao and Yuan, Yuhui and Gu, Shuyang and Chen, Bohan and Hang, Tiankai and Cheng, Mingxi and Li, Ji and Zheng, Liang},
journal={arXiv preprint arXiv:2406.04314},
year={2024}
}
SHA256
B6C2C16F3EF579885F10E94468D8F7196D09464002D116C115432207F4B1F8AB
ID Model ModelTrends.ai
#5399
Laporkan kandungan
Contoh render
Prompt dan tetapan yang dikongsi oleh orang yang membuat render ini di Civitai. Pilih render untuk melihat cara ia dibuat.
Tetapan
Sampler
Euler a
Langkah
40
Guidance
3.5
Seed
2101208291
Saiz
832x1216
Clip skip
2
Berat yang digunakan
0.6
Prompt
masterpiece, best quality, amazing quality, very aesthetic, absurdres, newest, volumetric lighting, holographic colors, cute, gradient colors, dark background, masterpiece, best quality, amazing quality, very aesthetic, absurdres, newest, volumetric lighting,1man,
BREAK
nsfw,
ambient occlusion, raytracing, soft lighting, blum effect, masterpiece, absolutely eye-catching, particles, glow, shining, intricate cinematic background, (beautiful face),
BREAK
masterpiece, highly detailed,detailed eyes, 1man, Loki, God of mischief, Loki god of mischief, detached virtual horns, Putting on a virtual holographic helmet with green glowing eyes and wielding 2 glowing green neon daggers while throwing several daggers with the other hand while covering his face in defense position , Shinobi shozoku, cyber ninja wear,, from front , detailed background, full body,dynamic ninja pose, , short hair, , closed mouth, ring, see-through, , gloves, ,, jewelry, cyberpunk shrine, shrine background, neon straight virtual cyber horns, wide sleeves, red eyes, lace, (glowing kunai:1.5), neon kunai, glowing, zPDXL3, detailxl, Score_PnyReal,
Prompt negatif
lowres, (worst quality,bad quality,low quality:1.2),
bad anatomy, old, bad hands, bad fingers, low quality eyes, bad eyes,
Tetapan
Guidance
4
Saiz
1595x2214
Prompt
1girl, fairy, (minigirl:1.5), solo, long hair, pointy ears, small breasts, delicate, translucent wings, insect wings, under mushroom, sad, big wings, dress, sleeves, floral print, translucent clothing, frilled dress, barefoot, hair ornament, covered nipples, wet clothes, sitting, looking up, (hugging own legs:1.1), hiding, wet, forest, mushroom, (rain:1.3), (water drop:1.3), fantasy, outside, melancholic, (monochrome:1.2)
Prompt negatif
loli, bad quality, worst quality, sketch, censor, censored, watermark, symmetry, plain simple background, (sparks, dust, noise:1.0), (blank eyes, empty eyes, simple eyes:0.8), 3d, realistic, photo,
Tetapan
Sampler
res_multistep_kl_optimal
Langkah
30
Guidance
5.5
Seed
620580403435191
Saiz
768x1344
Clip skip
2
Prompt
masterpiece, best quality, amazing quality, klskx, (nsfw, explicit:0.7), side view, 3/4 turned face, looking at viewer, 1girl, solo, red hair, long hair, detailed eyes, multicolored eyes, heterochromia, black sclera, red eyes, blue eyes, close-up, eyebrows, eyelashes, hooded cloak, decorated clothes, black feathered wings, big wings, standing before portal, (glowing portal:1.2), hand reaching out, swirling magic energy, ancient stone archway, runic symbols, dark ruins, intricate details, (blurry background, depth of field, bokeh:0.9) <lora:klskx_wai:0.9> <lora:spo_sdxl_10ep_4k-data_lora_webui:2.0>
Prompt negatif
child, loli, patreon logo, bad quality, worst quality, worst detail, sketch, censor, censored, (watermark:1.0), signature, blurry, plain simple background, blank eyes, messy hair, hydrokinesis, sparks, glares, lens flare,
Tetapan
Sampler
res_multistep_kl_optimal
Langkah
20
Guidance
5
Seed
537791972071645
Saiz
768x1280
Clip skip
2
Prompt
masterpiece, best quality, amazing quality, klskx, nsfw, explicit, 1girl, solo, pink hair, high ponytail, long ponytail, heterochromia, blue eyes, round glasses, pink eyebrows, eyelashes, mecha pilot suit, skin tight, cameltoe, outdoors, city view, intricate details, (blurry background, depth of field:1.1) <lora:klskx_wai:0.8> <lora:spo_sdxl_10ep_4k-data_lora_webui:5.0>
Prompt negatif
child, loli, patreon logo, bad quality, worst quality, worst detail, sketch, censor, censored, watermark, signature, blurry, plain simple background, blank eyes, messy hair, hydrokinesis, earrings, suction cups,
Tetapan
Sampler
DPM++ 2M Karras
Langkah
28
Guidance
5.5
Seed
1407126189
Saiz
832x1216
Clip skip
2
Berat yang digunakan
0.3
Prompt
scifi, masterpiece, best quality, great composition, glitching cyber ghost appears as a spectral, humanoid figure partially constructed of flickering digital elements. Its form constantly glitches and shifts, with fragments of its body dissolving into pixelated static or distorted neon lines. The skin may have a translucent, holographic quality, with glowing veins of circuit patterns running underneath. Its face might be a skull, glowing mask with shifting digital symbols or fragments of broken code flashing erratically. The figure could emanate an eerie light in shades of electric blue, neon purple, and stark white, contrasting against shadowy surroundings. Bits of data "noise" or fragmented particles might trail behind as it moves, as if it's perpetually disintegrating and reconstructing itself. The air around it crackles with faint digital distortion, and its presence feels both menacing and otherworldly, as though it exists somewhere between a corrupted program and a haunting spirit.
Prompt negatif
lowres, (worst quality,bad quality,low quality:1.2),
bad anatomy, old, bad hands, bad fingers, low quality eyes, bad eyes, horns, legs, female, male, human, girl, woman
Tetapan
Sampler
Euler
Langkah
30
Guidance
4.5
Seed
49716620367
Saiz
832x1536
Prompt
masterpiece, best quality, good quality, very aesthetic, absurdres, newest, 1girl, hanfu,see-through,
Sitting on the glowing crescent moon, the tranquil lake, reflections, and toes lightly touching the water <lora:spo_sdxl_10ep_4k-data_lora_webui:0.6> <lora:ponyv5_noobV065S_1_adamW-000013:0.7> <lora:USNR_STYLE_XL_lokr:0.45> <lora:GBF_Illustrious:0.4> <lora:748cmSDXL:0.8>
Prompt negatif
lowres, (worst quality,bad quality,low quality:1.2),
bad anatomy, old
Tetapan
Sampler
res_multistep_kl_optimal
Langkah
30
Guidance
4
Seed
906966760796694
Saiz
768x1280
Clip skip
2
Prompt
masterpiece, best quality, amazing quality, klskx, nsfw, explicit, 1girl, lisa \(genshin impact\), hat, seductive smile, playful expression, latex bodysuit, corset, (from side:0.9), standing on one leg, (cameltoe:0.9), v over eye, book in hand, indoors, (grand library:1.2), wooden bookshelves, scattered open books, dim candlelight, luxurious furniture, (dark shadows:0.9), (high contrast:0.9), volumetric lighting, intricate details, blurry background, depth of field <lora:spo_sdxl_10ep_4k-data_lora_webui:2.0> <lora:illustriousXL_stabilizer_v1.23:1.0> <lora:illustrious_quality_modifiers_masterpieces_v1:0.6> <lora:klskx_wai:0.7000000000000001>
Prompt negatif
loli, bad quality, worst quality, sketch, censored, watermark, signature, blurry, plain simple background, blank eyes, hydrokinesis, symmetry, huge breasts, dust particles in air, thighhighs,
Model serupa
Model AI untuk penjanaan imej, video dan teks
spo xl ialah LoRA untuk keluarga Stable Diffusion XL 1.0 yang disenaraikan di ModelTrends.ai, katalog baca sahaja untuk model AI imej, video dan teks sumber terbuka. Bandingkannya dengan model Stable Diffusion XL 1.0 lain, semak skor habanya untuk melihat trendnya, dan bukanya di PirateDiffusion atau BitVector untuk mencubanya.
Layari mengikut keluarga model
Model Text / LLM
411
Model Anima
509
Model Chroma
14
Model Flux
1,071
Model Flux 2 / Klein
1,237
Model MiniMax H3
331
Model Hunyuan
197
Model Ideogram
20
Model Krea2
674
Model Qwen
17
Model Qwen2
36
Model Ltx2
150
Model Stable Diffusion 1.5
6,465
Model Stable Diffusion XL 1.0
15,432
Model Zimage
804
Model Wan
919
Model
Trends
.ai
© ModelTrends.ai
|
Dibuat di Jepun
|
© 2026