Loading
Vous avez un workflow ComfyUI ? Obtenez une documentation lisible directement dans votre navigateur.Rendre un workflow lisible.Essayer maintenant ↗
Gilded detail, candlelit stone, and skin that reads like skin - without giving up anime structure or prompt adherence.
Source du modèle
Description de la source
RELIQUARY XL
Gilded detail, candlelit stone, and skin that reads like skin - without giving up anime structure or prompt adherence.
WHAT IT IS
• A semi-real Illustrious merge built around one idea - keep anime structure and tag adherence, but render it in real materials and real light
• It sits between flat cel anime and full realism. Faces and proportions stay illustrated, fabric and metal and stone do not
• Gold filigree reads as raised metal, lace reads as lace, and a cathedral behind the subject resolves into arches and candles instead of dissolving into blur
• Sibling model to VITRINE XL, which does the 3DCG collectible-figure look. Same family, opposite surfaces
Sources
1 sourceArtefacts du modèle
1 artefactreliquaryXL_v11.safetensors
safetensors · FP16 · 6,46 GB · SHA-256 3910977a76fa…e6fe · Civitai
TéléchargerExtraits de sources
1 extraitSource context: 262 downloads · Type Checkpoint · Base model Illustrious
WHAT IT DOES WELL
• Ornate materials - brocade, embroidery, filigree, gemstones, plate armour
• Atmospheric light - candlelight, volumetric god rays, rim light, wet neon
• Semi-real skin while keeping anime eye and face structure
• Environment depth - backgrounds hold detail instead of flattening
• Hands - measurably better than the Illustrious backbone it started from
• Character knowledge - Danbooru character tags survived the merge intact
WHAT IT IS NOT
• Not a photoreal model - if you want pores and real bone structure, this is the wrong end of the spectrum
• Text and signage are unreliable, as with every SDXL model. Expect decorative gibberish on shop signs and neon
• Complex hand interactions (both hands gripping something) are hit and miss. Framing tighter so the hands get more pixels helps more than any negative prompt
RECOMMENDED SETTINGS
• Sampler - Euler a
• Steps - 28
• CFG - 5.0 (4 to 6 works)
• Clip skip - 2
• Resolution - 832x1216 portrait, 1216x832 landscape
• VAE - baked in, do not load an external one
• Hires - 1.5x latent upscale, denoise 0.45, 16 steps
• No trigger words. Danbooru tags
QUALITY PREFIX
Lead with this - it matters on this merge:
masterpiece, best quality, amazing quality, very aesthetic, absurdres
SUGGESTED NEGATIVE
bad quality, worst quality, worst detail, sketch, censored, watermark, signature, jpeg artifacts, bad anatomy, bad hands, extra digits, fewer digits, missing fingers, blurry, lowres, text, logo
PROMPTING NOTES
• Name your materials - ornate black and gold gown, lace high collar, ruby brooch gets far more out of this model than pretty dress. Material nouns are where the merge earns its keep
• Name your light - candlelight, volumetric god rays, rim lighting and reflected neon all land distinctly
• Always put logo in your negative for anything with a cup, bottle, sign or product. This model reproduces a certain green circular coffee logo with some enthusiasm - it turned up unprompted in two separate cafe scenes during testing. Fine for personal use, a trademark problem if you sell the output
• Do not push CFG past about 7 - it hardens the shading and undoes the semi-real look
• Under-specify known characters - use the character tag and describe outfit and scene only. Adding hair or eye colour overrides the character tag and you get the wrong character
