AI Image Negative Prompt Ignored: How to Fix It

Q: What is the difference between `guidance_scale` and `true_cfg_scale` on Flux?

`guidance_scale` is Flux's distilled guidance value (typically 3-4) and does not use a negative prompt. `true_cfg_scale` is real classifier-free guidance: it only kicks in above 1, and only then is your `negative_prompt` actually applied. If you set a negative prompt but leave `true_cfg_scale` at 1, it is silently discarded.

Listed 'text, watermark, extra fingers' in the negative prompt and got all three anyway? It is usually one of seven causes: not wired up, CFG too low, contradicted by the positive prompt, truncated past 75 tokens, wrong weight syntax, an overtrained checkpoint, or a Flux model that ignores negatives by design.

Published: May 24, 2026 Updated: Jun 17, 2026 Author: AI Productivity Guide Team 🌐 查看中文版本

You painstakingly listed blurry, low quality, text, watermark, extra fingers, deformed hands, jpeg artifacts in the negative prompt. The image came back with text on a t-shirt, a faint watermark in the corner, and six fingers. It looks like the negative prompt is being thrown away.

Fastest fix: log the request body and confirm the model actually received a non-empty negative_prompt, then raise CFG to 8 and re-run. Those two checks resolve the majority of “negative prompt ignored” cases. If you are on a Flux model (Flux Dev, Flux Schnell, or Flux.2), the negative prompt is disabled by design unless you enable true CFG — jump to the Flux section.

Sometimes the prompt literally never arrives — the client did not pass it. More often the model is parsing it fine, but the concept you are negating is baked too strongly into the checkpoint weights or summoned by your own positive prompt. Treat the negative prompt as one lever among several, not a magic eraser.

Which bucket are you in?

Symptom	Likely cause	Go to
Works in the web UI but not your script	Field not sent / wrong field name	Cause 1, Step 1
Unwanted thing fades at high CFG	CFG too low	Cause 2, Step 2
Positive prompt describes a magazine, poster, sign	Positive prompt summons it	Cause 3, Step 3
Only the last items in a long negative fail	Past the 75-token cutoff	Cause 5, Step 4
`(word:1.4)` shows as literal text	Wrong weight syntax for the runtime	Cause 6, Step 5
Watermark survives even an empty positive prompt	Overtrained checkpoint	Cause 4, Step 6
Clean base image, dirty after refiner	Refiner not given the negative	Cause 7, Step 7
Model is Flux (Dev/Schnell/.2)	Negatives disabled by default	Step 8

Common causes

Listed by frequency in real pipelines.

1. Negative prompt never made it to the model

In API calls this is the most common — the field is named negative_prompt, neg_prompt, or unconditional_prompt depending on the provider. A typo means an empty string is sent. UIs sometimes silently drop the field on certain workflows (e.g. img2img in some Comfy nodes).

How to spot it: Echo the outgoing request body. If negative_prompt is missing or empty, the model never saw it.

2. CFG scale is too low for the negative prompt to bite

CFG controls how strongly the model pushes toward positive and away from negative. At CFG 4-5 the negative prompt has weak effect; at 7-9 it bites; above 12 it can over-correct into ugly artifacts.

How to spot it: Re-run at CFG 8 with negative prompt unchanged. If the unwanted element disappears, low CFG was the cause.

3. Positive prompt is summoning what the negative is rejecting

If the positive prompt mentions “magazine cover” the model will produce text — and “text” in the negative is fighting the entire positive prompt. Negative prompts cannot overcome a contradiction inside the positive prompt.

How to spot it: Read the positive prompt and ask “would a literal-minded model add the thing I am trying to remove?” If yes, fix the positive prompt first.

4. Model has the unwanted concept overtrained

Some checkpoints (especially photorealism fine-tunes) are saturated with watermarks because their training set included stock photos. The negative prompt nudges but does not overcome a hard-coded bias.

How to spot it: Generate with a fully empty positive prompt and full negative prompt. If the watermark still appears, the base model is the source.

5. Negative prompt is too long and got truncated

CLIP’s hard limit is 77 tokens, but two of those are reserved for the start/end markers, so you get 75 real tokens (roughly 50-60 words). In raw Diffusers, anything past that is silently dropped — sometimes including the very concept you cared about most. AUTOMATIC1111, Forge, Fooocus, InvokeAI, and ComfyUI work around this by splitting the prompt into 75-token chunks (so the effective limit climbs to 150, 225, and so on), but a plain pipe(...) call in a script does not.

How to spot it: Count tokens with a CLIP tokenizer. If your negative prompt runs past 75 tokens and you are calling Diffusers directly, the tail is being ignored. In Diffusers, watch the console for the warning The following part of your input was truncated because CLIP can only handle sequences up to 77 tokens.

6. Token weight syntax is wrong for the platform

(blurry:1.4) works in AUTOMATIC1111 but not in raw Diffusers. [blurry:1.4] is interpreted differently in different UIs. Wrong syntax means the parser sees garbled text and the weight is not applied.

How to spot it: Test with plain words first, no parentheses or weights. If plain words work and weighted versions do not, syntax is the issue.

7. SDXL refiner pass re-introduces what the base removed

SDXL’s two-stage pipeline runs a base model then a refiner. If the negative prompt is only passed to the base, the refiner can re-add detail (including watermark-like noise) on top of a clean base image.

How to spot it: Run base-only (refiner disabled) and compare. If base-only is clean and refined is dirty, the refiner pass is at fault.

8. You are on a Flux model, which ignores negatives by design

Flux Dev, Flux Schnell, and Flux.2 are guidance-distilled: the classifier-free guidance behavior is baked into the weights, so they run at an internal CFG of 1 where a negative prompt has no effect. As of June 2026 this is the single most common reason a negative prompt “does nothing” for people who recently switched off SDXL. The negative field exists in the UI but is inert unless you turn on true CFG.

How to spot it: Check your checkpoint name. If it contains flux and your negative prompt has zero effect at any CFG, this is why. See Step 8 for the fix.

Before you start

Save one “bad” output and note the exact positive + negative prompt strings and all parameters.
Confirm the API or UI you are using supports negative prompts at all (some basic providers do not).
Decide which unwanted element is most important to remove — fixing one at a time is faster than fighting all at once.

Information to collect

Full positive prompt and negative prompt strings (copy from request log, not from memory).
CFG scale, sampler, model name, model version.
Token count of both prompts (use a CLIP tokenizer).
Whether you are running base only, refiner only, or both.
The exact field name your client uses (negative_prompt vs neg_prompt vs other).
The model family (SDXL, SD 1.5, Flux Dev/Schnell/.2, Qwen-Image) — this decides whether negatives even apply.

Step-by-step fix

Ordered by ROI.

Step 1: Confirm the negative prompt is actually sent

Log the outgoing request body:

import json
print(json.dumps(payload, indent=2))

Or in curl:

curl -v -X POST $API_URL \
  -H "Authorization: Bearer $KEY" \
  -d '{"prompt": "...", "negative_prompt": "blurry, text, watermark", "cfg_scale": 7.5}'

If the negative_prompt key is absent, fix the field name. Half of “negative prompt ignored” tickets end here.

Step 2: Raise CFG and re-test

Try CFG 8.0, then 9.5. Watch for over-correction (washed-out colors, fried-looking edges) but if CFG 8 makes the unwanted element disappear, low CFG was the cause.

{"cfg_scale": 8.5}

Step 3: Fix the positive prompt to stop summoning the unwanted thing

If you wrote “studio portrait, magazine cover quality, fashion editorial style” and negative-prompted “no text”, you are contradicting yourself. Rewrite:

positive: studio portrait, soft rim light, neutral grey backdrop, plain background
negative: text, watermark, signature, letters, logo

A clean positive prompt does more work than a frantic negative prompt.

Step 4: Trim the negative prompt to under 60 tokens

Order matters: put the most important rejections first, because in raw Diffusers anything past token 75 is dropped, and even inside the limit the first tokens carry the most weight.

negative: text, watermark, extra fingers, deformed hands, blurry, low quality

Drop low-value generic words like “ugly”, “bad”, “weird” — they rarely help and crowd out specific tokens.

If you genuinely need a long negative in AUTOMATIC1111, Forge, or ComfyUI, use the uppercase BREAK keyword to start a fresh 75-token chunk and group related rejections together, for example text, watermark BREAK extra fingers, deformed hands. BREAK must be uppercase; lowercase break is treated as an ordinary word.

Step 5: Use platform-correct weighting syntax

For AUTOMATIC1111 / Forge / Fooocus / ComfyUI / InvokeAI, the (word:1.5) syntax all works the same way:

(watermark:1.5), (text:1.4), (extra fingers:1.3), blurry, deformed

For raw Diffusers, that parenthesis syntax is just literal text — weighting requires the compel library, which uses its own word++ / word-- syntax and feeds the result in via negative_prompt_embeds. Without it, parens do nothing.

For SD 1.5 and SDXL specifically, a textual-inversion embedding is often stronger than typed words: drop the EasyNegative embedding into your embeddings folder and add EasyNegative to the negative prompt. One token encodes a bundle of “bad quality / bad anatomy / artifact” concepts that plain words struggle to express.

Step 6: Switch checkpoint if the concept is overtrained

If a stock-photo-trained checkpoint always emits watermarks, swap to one trained on cleaner data (a “photo-real-clean” variant, or the base SDXL without stock fine-tunes). Related: AI image watermark residue.

Step 7: Pass the negative prompt to both base and refiner (SDXL)

In SDXL pipelines, ensure the refiner stage also receives the negative prompt:

image = pipe(
    prompt=pos,
    negative_prompt=neg,
).images[0]
image = refiner(
    prompt=pos,
    negative_prompt=neg,
    image=image,
).images[0]

Many tutorials forget the refiner call’s negative_prompt argument — leaving it default re-introduces the bad content.

Step 8: Flux needs true CFG or Dynamic Thresholding

If your model is Flux Dev, Flux Schnell, or Flux.2, a plain negative prompt has no effect because guidance is distilled into the weights. You have three options, in order of preference as of June 2026:

In Diffusers, use true CFG. The Flux pipeline supports negative_prompt only when true_cfg_scale is greater than 1. This is the cleanest, officially-supported path:
```
image = pipe(
    prompt=pos,
    negative_prompt="blurry, low quality, text, watermark",
    true_cfg_scale=3.5,   # negative is IGNORED if this stays at 1
    num_inference_steps=50,
).images[0]
```
Note true_cfg_scale is a separate parameter from the Flux guidance_scale (the distilled guidance, usually 3-4). Leaving true_cfg_scale at its default of 1 silently discards the negative prompt. See the Diffusers Flux pipeline docs for the exact parameter list.
In ComfyUI, install sd-dynamic-thresholding and add the DynamicThresholdingFull node between your Flux model and the KSampler. This lets you push CFG above 1 and pass a real negative prompt. Results are hit or miss and it roughly doubles render time, but it works.
Skip negatives entirely and prompt positively. This is often the better Flux workflow: instead of negating “extra fingers”, describe the correct outcome — a hand with exactly five fingers, anatomically correct. Flux follows detailed positive prompts well, which is why it leads on prompt adherence in 2026.

Verify

Run a 4-image grid with the fixed prompts. The unwanted element should be absent or rare across all four.
Generate with negative prompt empty and confirm the unwanted element returns. Proves the negative prompt is now load-bearing.
Generate one extreme case (CFG 12, max weights) and confirm the model can fully suppress the element — even if you would not ship at that CFG, it confirms the lever works.

Long-term prevention

Keep a tested negative prompt library per checkpoint. What works on SDXL base does not work on Flux.
Log positive + negative + CFG + sampler + checkpoint together. When negative prompt fails, you want to bisect from a known-good config.
Fix the positive prompt first. Negative prompt is a corrector, not a rewriter.
Cap negative prompt at 60 tokens; put critical rejections in the first 20.
Validate the request payload in CI for your custom pipelines — assert that negative_prompt is present and non-empty, and that true_cfg_scale is greater than 1 on Flux jobs.
Re-test your negative prompt library every time you change checkpoint version.
Record which model family each profile targets. SDXL negatives are the most powerful, Flux needs true CFG or positive framing, and Qwen-Image (Alibaba, Apache 2.0) behaves differently again — do not assume one library transfers across families.

Common pitfalls

Treating the negative prompt as a wishlist and stuffing 200 words in it. The tail is dropped.
Writing “no text” in the negative prompt — model parses “no” and “text” as separate concepts; “text” is what registers. Just write the concept you reject.
Using AUTOMATIC1111 weight syntax in a raw Diffusers script that does not parse it.
Forgetting that hosted providers may strip or remap the field. Check their docs.
Raising CFG to 15+ to brute-force suppression — output gets fried and saturated. Better to fix the positive prompt or change checkpoint.
Using a negative prompt on Flux without setting true_cfg_scale greater than 1 (Diffusers) or adding Dynamic Thresholding (ComfyUI). The field is inert otherwise.
Assuming negative prompt fixes everything. Use reference image or ControlNet for structural problems negative prompts cannot fix.

FAQ

Q: Should I always use a long negative prompt by default?

No. Default negative prompts copied from forums often contain contradictions or waste tokens. Use a short, targeted negative prompt focused on the actual problems your checkpoint shows.

Q: My positive prompt is short and clean and the model still adds text. What now?

Test with a different checkpoint. Some realism fine-tunes have watermark/text deeply baked in and no negative prompt fully removes them.

Q: Can I use a regex or template for negative prompts?

Treat the negative prompt like a checklist: per-checkpoint, per-style. There is no universal regex. Save 3-4 named profiles (“portrait-clean”, “product-clean”, “anime-clean”) and reuse.

Q: Does Flux respect negative prompts the same way?

No. Flux Dev, Flux Schnell, and Flux.2 are guidance-distilled and ignore negative prompts at their default settings. In Diffusers you re-enable them by setting true_cfg_scale above 1; in ComfyUI you add the DynamicThresholdingFull node from the sd-dynamic-thresholding extension. Both cost extra render time, so for Flux many people skip negatives and describe the desired result positively instead.

Q: What is the difference between guidance_scale and true_cfg_scale on Flux?

guidance_scale is Flux’s distilled guidance value (typically 3-4) and does not use a negative prompt. true_cfg_scale is real classifier-free guidance: it only kicks in above 1, and only then is your negative_prompt actually applied. If you set a negative prompt but leave true_cfg_scale at 1, it is silently discarded.

Q: Which checkpoint should I switch to if mine keeps emitting watermarks?

For the deepest library of clean community fine-tunes and the strongest negative-prompt response, stay on SDXL. If you need legible in-image text or commercial licensing, Qwen-Image (Apache 2.0) is a strong 2026 option. Flux leads on prompt adherence but, as above, treats negatives differently.

Tags: #Troubleshooting #ai-image #negative-prompt #Prompt engineering #diffusion