Differences

This shows you the differences between two versions of the page.

--- ai:atlas-prompts [2026/04/24 14:05] – [🛡️ Fantasy Warrior] mh
+++ ai:atlas-prompts [2026/07/15 22:43] (current) – [Workflow] mh
@@ Line 615: / Line 615: @@
 ----
+===== Local AI add-ons =====
+==== Positive Baseline (SDXL) ====
+<code txt>
+high quality image, detailed, coherent composition, <REST OF PROMPT>
+</code>
+==== Negative Baseline ====
+<code txt>
+low quality, worst quality, low resolution, blurry, jpeg artifacts, watermark, signature, logo, unwanted text, text artifacts, bad anatomy, distorted anatomy, deformed body, extra limbs, missing limbs, extra fingers, missing fingers, fused fingers, malformed hands, malformed feet, duplicate subject, cropped subject, out of frame
+</code>
+==== Workflow ====
+Lock the whole ritual circle:
+<code txt>
+same prompt
+same negative prompt
+same checkpoint type family
+same resolution
+same seed policy
+same sampler
+same scheduler
+same steps
+same CFG
+same VAE behavior
+no LoRA
+no ControlNet
+no IPAdapter
+no ref image
+no upscale
+no inpainting
+</code>
+At first: checkpoints only.
+Then later:
+<code txt>
+checkpoint + LoRA
+checkpoint + sampler variation
+checkpoint + CFG variation
+checkpoint + style LoRA
+checkpoint + ControlNet
+</code>
+Don’t mix those into the baseline.
+Baseline settings suggestion for SDXL checkpoints:
+<code txt>
+Resolution: 1024 x 1024 for square prompts
+Steps: 30
+CFG: 6
+Sampler: dpmpp_2m
+Scheduler: karras
+Batch size: 1
+Seed: fixed per prompt across all checkpoints
+LoRA: none
+Refiner: none, unless every checkpoint test uses the same refiner strategy
+</code>
+For non-square prompts, use SDXL-friendly sizes close to 1 megapixel:
+<code txt>
+Portrait: 832 x 1216
+Landscape: 1216 x 832
+Wide: 1344 x 768
+Square: 1024 x 1024
+</code>