Prompts that work on Promptchan.ai

Prompt quality collapses when users stack adjectives. Forcing three distinct axes (subject / style / light) plus one mood anchor is enough to produce consistent results without overwhelming the diffusion model.

Diffusion models don't naturally maintain character consistency across generations. Listing 2-3 immutable features each time forces the model to anchor on them. More than 3 features and you get drift; fewer than 2 and the character morphs.

Negative prompts are more effective than positive prompts for fixing common generation failures. Hands are the #1 failure mode. The model actively needs to be told to not mess them up. Overloading negatives degrades the positive prompt's influence.

Models are trained on captioned images. Specific photography / style references tokenize better than adjective stacks. One good reference phrase shifts the entire output more reliably than 'cinematic, moody, atmospheric, dramatic'.

Diffusion models generate poses from captioned photo datasets, which describe bodies in concrete part-and-position language. 'Leaning against a doorframe' generates reliably; 'feeling relaxed' doesn't. Body-part anchors (shoulder, elbow, knee) act like keypoints the model latches onto.

Outfit drift is the second-biggest consistency failure after face drift. Listing items as color-plus-material tokens ('black cotton t-shirt') rather than style words ('casual shirt') locks the generation to specific visual features the model can reproduce across seeds.

Emotion words ('happy', 'sad') map to averaged expressions in training data and produce bland results. Describing mouth and eyes as independent features activates more specific regions of the model's expression space. This is how professional prompt engineers direct faces.

Diffusion models render environments well when given the light-source / surface / background-object triad because that matches how scene captions are typically structured in training data. Vague settings ('a cozy room') produce averaged slop; the triad produces specific, coherent scenes.

Anime diffusion models (most Waifu Diffusion / NovelAI descendants) were trained on Danbooru/Gelbooru tags, not natural language. Booru syntax (tag-based, comma-separated, no 'a'/'the') produces sharper outputs than prose because it matches the training caption format exactly.

Photo-realism degrades when you mix photography terms with art terms: the model averages between modes. Sticking strictly to camera/lens/light/depth-of-field vocabulary and explicitly including 'skin texture visible' counteracts the smoothing bias most SDXL-era models have.

'Golden hour' alone triggers a generic warm-orange preset. Specifying light direction (camera-left), affected surfaces (hair, jawline), shadow behavior, and atmosphere (haze) forces the model to render the physics rather than the stereotype. This matches how cinematographers actually describe light.

Single-artist references produce stronger stylistic coherence than stacked adjectives because the model has dense clusters of captioned work per artist. Stacking multiple artists averages them into mush. The decade/genre/medium fallback works for lesser-known aesthetics and is safer than guessing at name recognition.

Full-body shots distribute pixel budget across the whole figure, which is why faces look melted at distance. Explicit close framing with 'face fills 60% of frame' and 'sharp focus on eyes' concentrates the model's rendering attention where it matters for character identity.

Diffusion models render full-body figures poorly because they don't naturally model weight distribution. Explicitly specifying which leg bears weight and where the other leg is positioned constrains the pose to a physically plausible one. This single instruction removes most of the uncanny-stance failures.