Negative prompts are dead — here's what actually works in SD3, Flux, and MJ V8.1

Three tools you use every day have quietly stopped responding to negative prompts the way you expect. The fix is different for each — and it's not complicated once you know which knob to turn.

SD3 and SD3.5: stop using `negative_prompt` entirely

Stable Diffusion 3 and SD3.5 were not trained with negative prompts. 1 Passing a negative_prompt string doesn't filter out unwanted elements — it injects noise into the conditioning signal, which behaves like switching to a different seed.

As fofr at Replicate puts it:

"Your negative prompt will not remove the elements you don't want; instead, it will introduce noise to your conditioning and simply vary your output, kind of like using a different seed."

The ComfyUI team arrived at the same conclusion from their own testing: SD3 performs best when the negative conditioning is zeroed out entirely. 2 The official Diffusers examples for both SD3 Medium and SD3.5 either pass negative_prompt="" or omit the parameter altogether. 3

The replacement strategy: positive reframing. SD3 supports prompts up to 10,000 characters — far beyond SDXL's 77-token CLIP limit — so you have room to describe what you want in full sentences. 1 Here's the translation table:

What you'd have put in `negative_prompt`	Replace with this in your main prompt
`no bad anatomy, extra limbs`	`anatomically correct proportions, exactly five fingers on each hand`
`no blur, no artifacts`	`sharp focus, crisp details, clean edges`
`no cartoon, no anime`	`photorealistic, DSLR photograph, raw photo`
`no bad face, deformed face`	`symmetrical features, natural skin texture, beautiful face`
`no watermark, no text`	`clean image, no overlaid text or symbols`

CFG is your secondary lever. SD3 uses rectified flow — CFG behaves differently than in SD1.5/SDXL, acting more as a prompt-adherence dial than a content filter. Keep it low:

SD3.5 Large: CFG 3.5 (model card recommendation) 4
SD3.5 Medium: CFG 4.5, 40 steps 5
SD3 Medium default: 7.0 — if outputs look oversaturated or "burnt," lower to 3.5–4.5

For anatomy specifically — SD3.5 Medium only: enable Skip Layer Guidance (SLG). The SD3.5 Medium model card explicitly recommends it: "We recommend sampling with Skip Layer Guidance for better structure and anatomy coherency." 5 SLG is a ComfyUI technique that skips specific transformer layers during guidance, replacing what negative prompts used to do for bad anatomy, extra limbs in SDXL. In the official inference script, pass --skip_layer_cfg True. 6

SD3 vs SDXL book cover comparison — SD3 renders text correctly while SDXL produces garbled characters — SD3 rendering the same children's book cover prompts as SDXL — text legibility is one of the places where the long positive-only prompt strategy pays off clearly. 1

Flux: no `negative_prompt` field exists anywhere in the API

Every Flux endpoint on fal.ai was audited on 2026-05-28. Result: zero native negative_prompt fields across all seven active endpoints — Flux.1 [dev], Flux.1 [schnell], Flux Pro v1.1, Flux Pro v1.1 Ultra, Flux 2 Pro, Flux 2 Dev, and Flux 2 Flex. [[cite:7|FLUX.1 [dev] API Docs – fal.ai|[https://fal.ai/models/fal-ai/flux/dev/api]]](https://fal.ai/models/fal-ai/flux/dev/api]]) [[cite:8|FLUX1.1 [pro] ultra API Docs – fal.ai|[https://fal.ai/models/fal-ai/flux-pro/v1.1-ultra/api]]](https://fal.ai/models/fal-ai/flux-pro/v1.1-ultra/api]])

One historical fal.ai marketing page for Flux dev mentioned negative prompts in its copy — but the actual API schema has no such field. It was copied from SDXL documentation, not a hidden feature.

Flux.1 uses T5-XXL as its text encoder; Flux.2 uses a Mistral-3 24B VLM. Neither architecture processes prompts as tokenized tags — they read natural language prose. The (tag:1.2) weight syntax from Stable Diffusion is meaningless here. 7

SD3 Explorer interface showing a long natural-language prompt on the left alongside its generated portrait output on the right — The SD3 Explorer demo — the left panel holds a multi-sentence positive prompt, the right is the output. No `negative_prompt` field in sight. 1

Your four replacement strategies, by model:

Strategy 1 — Positive description (works on all Flux models). Write what you want, precisely. no blur → sharp focus, tack-sharp details. no extra fingers → anatomically correct hands with exactly five fingers. The T5/Mistral encoder understands full sentences — use them.

Strategy 2 — Raise guidance_scale (Flux dev, Flux 2 Dev, Flux 2 Flex only). Higher guidance_scale makes the model stick more closely to your prompt, which tightens the gap left by the missing negative channel. 8 9

Model	Default `guidance_scale`	Suggested range to try
Flux.1 [dev]	3.5	5–7
Flux 2 Dev	2.5 (range 0–20)	4–6
Flux 2 Flex	3.5 (range 1.5–10)	5–8

Strategy 3 — enable_prompt_expansion (Flux 2 Dev only). Set enable_prompt_expansion: true. The model automatically expands a short prompt into a richer description, which improves both output quality and adherence. Useful when your prompt is concise and the model keeps adding unwanted elements.

Strategy 4 — enhance_prompt and raw (Flux Pro v1.1 / Ultra only). Flux Pro v1.1 Ultra has no guidance_scale and no num_inference_steps at all — it's the most locked-down endpoint in the family. Your only real levers are:

enhance_prompt: true — server-side prompt enrichment before generation
raw: true — "generate less processed, more natural-looking images" [[cite:8|FLUX1.1 [pro] ultra API Docs – fal.ai|[https://fal.ai/models/fal-ai/flux-pro/v1.1-ultra/api]]](https://fal.ai/models/fal-ai/flux-pro/v1.1-ultra/api]])

For Ultra: write the best positive prompt you can, turn on enhance_prompt, and use raw: true if the model is over-beautifying.

MJ V8.1: `--no` works, `::-1` probably doesn't

The Midjourney situation is genuinely confusing because the official documentation has an internal contradiction. Here's what the evidence actually shows.

--no is confirmed supported in V8.1. The official Version compatibility table shows a ✅ for --no in V8.1. 10 Mechanically, --no red is equivalent to giving red a weight of -0.5 in a multi-prompt. 11

Multi-Prompting (:: syntax) is not supported in V8.1. The same Version page marks Multi-Prompting as ❌ for both V7 and V8.1. The Multi-Prompts documentation explicitly lists supported versions as "1, 2, 3, 4, Niji 4, 5, Niji 5, 6, Niji 6, and 6.1" — V7 and V8.1 are absent. 11 Since ::-1 is part of the multi-prompt syntax, it likely cannot be used in V8.1 at all.

Note: third-party guide Blake Crosley marks --no as also unsupported in V8.1, which directly contradicts the official compatibility table. 12 V8.1 is still in Alpha — treat both the --no status and the ::-1 limitation as subject to change and verify in-session.

MJ V8.1 compatibility diagram — --no parameter shows green checkmark, :: multi-prompt syntax shows red X indicating unsupported in V8.1 Alpha — V8.1 Alpha status at a glance: `--no` is in, `::` multi-prompt (and therefore `::-1`) is out. AI-generated illustration.

Practical --no strings that hold up in V6/V7 (and likely V8.1) per Blake Crosley: 12

# Keep outputs photorealistic
--no anime, cartoon, illustration, painting, drawing

# Remove graphic elements
--no text, watermark, signature, frame, border

# Reduce oversaturation
--no oversaturated, HDR, artificial

# Cinematic realism
--no anime, cartoon, illustration, painting, drawing, sketch, CGI, 3D render

Blake's operational rule: keep the list to 3–5 items, use specific terms, and only list things that have actually appeared in your outputs. Vague entries like bad or ugly waste slots. The total weight rule still applies — all weights in a prompt must sum to a positive number. still life painting:: fruit::-0.5 works (sum = 0.5); still life painting:: fruit::-2 errors (sum = -1). 11

--raw complements --no in V8.1. Raw mode turns off Midjourney's aesthetic auto-pilot, producing less stylized, more photographic output. 13 When you're fighting unwanted stylization, --raw reduces what --no would need to suppress.

Cross-tool cheat sheet

Tool / Model	`negative_prompt` field?	Primary alternative	Secondary lever	Copy-paste note
SD3 Medium	❌ adds noise	Positive reframe in main prompt	CFG 3.5–4.5	28 steps, dpmpp_2m + sgm_uniform
SD3.5 Large	❌ adds noise	Positive reframe in main prompt	CFG 3.5	Omit `negative_prompt` entirely
SD3.5 Medium	❌ adds noise	Positive reframe + SLG	CFG 4.5, 40 steps	`--skip_layer_cfg True` in CLI
Flux.1 [dev]	❌ no field	Positive description	guidance_scale 5–7	Default 3.5; raise to tighten adherence
Flux.1 [schnell]	❌ no field	Positive description	None (1–4 steps only)	Speed model; limited steering
Flux Pro v1.1	❌ no field	Positive description	`enhance_prompt: true`	No guidance_scale in base t2i
Flux Pro v1.1 Ultra	❌ no field	Positive description	`raw: true`	No guidance_scale at all
Flux 2 Dev	❌ no field	Positive description	guidance_scale 4–6	Default 2.5; `enable_prompt_expansion` available
Flux 2 Flex	❌ no field	Positive description	guidance_scale 5–8	Default 3.5; range 1.5–10
MJ V8.1	`--no` ✅ (Alpha)	`--no [terms]`	`--raw` (reduces stylization)	`::` multi-prompt ❌ in V8.1; keep `--no` list to 3–5 items

The pattern is consistent across all three ecosystems: describe the image you want, not the image you're avoiding. SD3 gives you sentence length to work with. Flux gives you guidance_scale to amplify adherence. Midjourney gives you --no for direct exclusion with a tight list. Use the right lever for the tool — and stop putting negative strings into fields that were never trained to read them.

Cover image: AI-generated illustration

Negative prompts are dead — here’s what actually works in SD3, Flux, and MJ V8.1

Negative prompts are dead — here's what actually works in SD3, Flux, and MJ V8.1

SD3 and SD3.5: stop using `negative_prompt` entirely

Flux: no `negative_prompt` field exists anywhere in the API

MJ V8.1: `--no` works, `::-1` probably doesn't

Cross-tool cheat sheet

참고 출처

Negative prompts are dead — here’s what actually works in SD3, Flux, and MJ V8.1

Negative prompts are dead — here's what actually works in SD3, Flux, and MJ V8.1

SD3 and SD3.5: stop using negative_prompt entirely

Flux: no negative_prompt field exists anywhere in the API

MJ V8.1: --no works, ::-1 probably doesn't

Cross-tool cheat sheet

참고 출처

SD3 and SD3.5: stop using `negative_prompt` entirely

Flux: no `negative_prompt` field exists anywhere in the API

MJ V8.1: `--no` works, `::-1` probably doesn't