haofanwang
commented
8 months ago It should be related to the prompt or weight scale. We don't tune the parameters carefully. Anyway, talk is easy, we will show you the code soon.
Open renderless opened 8 months ago
haofanwang
commented
8 months ago It should be related to the prompt or weight scale. We don't tune the parameters carefully. Anyway, talk is easy, we will show you the code soon.
wangqixun
commented
8 months ago In technical report, Fig. 6, "Jackie Chan" does not looks like reference image especially his nose. I supposed antelopev2 model should able to extract Jackie Chan very well as his images should be in Glint360k training data. Is this the limitation of face id encoder?
Also, the picture quality seems over saturated. Is it because of SDXL base model or your prompt?
It may be a problem with prompt. I changed the prompt and base model to generate a new image. At the same time, our model is constantly being optimized.
lucasjinreal
commented
8 months ago @wangqixun which model were using here?
wangqixun
commented
8 months ago @wangqixun which model were using here?
base model = https://civitai.com/models/43977?modelVersionId=227916
prompt = "cinema 4d render, high contrast, vibrant and saturated, sico style, dark and moody close-up shot of a handsome Saint-Pierrais man with a tired expression, (renaissance theme:1.1), colorful northern warrior, (glowing eyes:1.05), dynamic pose, hooded robe, surrounded by magical glow, floating ice shards, snow crystals, cold, windy background, frozen natural landscape in background cinematic atmosphere, highly detailed, sharp focus, intricate design, 3d, unreal engine, octane render, CG best quality, highres, photorealistic, dramatic lighting, artstation, concept art, cinematic, epic Steven Spielberg movie still, sharp focus, smoke, sparks, art by pascal blanche and greg rutkowski and repin, trending on artstation, hyperrealism painting, detailed character design, matte painting, 4k resolution"
neg prompt = "asian, (worst quality, low quality, thumbnail:1.4), signature, artist name, web address, cropped, jpeg artifacts, watermark, username, collage, grid, nude, topless, nsfw, naked, nipples"
The style is not very stable. Generate the image again.
wangqixun
commented
7 months ago @wangqixun which model were using here?
base model = https://civitai.com/models/43977?modelVersionId=227916
prompt = "cinema 4d render, high contrast, vibrant and saturated, sico style, dark and moody close-up shot of a handsome Saint-Pierrais man with a tired expression, (renaissance theme:1.1), colorful northern warrior, (glowing eyes:1.05), dynamic pose, hooded robe, surrounded by magical glow, floating ice shards, snow crystals, cold, windy background, frozen natural landscape in background cinematic atmosphere, highly detailed, sharp focus, intricate design, 3d, unreal engine, octane render, CG best quality, highres, photorealistic, dramatic lighting, artstation, concept art, cinematic, epic Steven Spielberg movie still, sharp focus, smoke, sparks, art by pascal blanche and greg rutkowski and repin, trending on artstation, hyperrealism painting, detailed character design, matte painting, 4k resolution"
neg prompt = "asian, (worst quality, low quality, thumbnail:1.4), signature, artist name, web address, cropped, jpeg artifacts, watermark, username, collage, grid, nude, topless, nsfw, naked, nipples"
The style is not very stable. Generate the image again.
Correct it
base model = https://civitai.com/models/84040?modelVersionId=196039
In technical report, Fig. 6, "Jackie Chan" does not looks like reference image especially his nose. I supposed antelopev2 model should able to extract Jackie Chan very well as his images should be in Glint360k training data. Is this the limitation of face id encoder?
Also, the picture quality seems over saturated. Is it because of SDXL base model or your prompt?