Denoising Strength in Image-to-Image: The Per-Tool Reference (2026)

TL;DR: Denoising strength sets how far an image-to-image run may travel from your input. In Stable Diffusion WebUI, ComfyUI and diffusers, higher means more change. In Midjourney, Ideogram, FLUX Ultra and Stability's control endpoints, higher means less change. The number is meaningless until you know which convention you are in. Checked August 29, 2026.

What is denoising strength in image-to-image?

It is the dial that decides how much of your input image survives.

The clearest published description sits in the reference implementation rather than any vendor's marketing. Hugging Face's diffusers library documents its strength argument as: "Indicates extent to transform the reference image. Must be between 0 and 1." Then the mechanism: "image is used as a starting point and more noise is added the higher the strength." At the ceiling, "A value of 1 essentially ignores image."

Image-to-image starts from your picture instead of pure noise, adds a controlled amount of noise, and denoises back down. Add a little and the model has almost no room. Add a lot and it has nothing left to hold.

So the parameter is always the same trade: fidelity to the input against freedom to follow the prompt. Every tool below implements it. None of them agree on what to call it or which way round to write the number.

Why does the same value do opposite things in different tools?

Because half the industry names the axis after the change and half names it after the image.

Stability's own API carries both conventions. On its v1 image-to-image endpoint, image_strength is documented as "How much influence the init_image has on the diffusion process. Values close to 1 will yield images very similar to the init_image while values close to 0 will yield images wildly different than the init_image." That endpoint also accepts step_schedule_start, and Stability states the relationship outright: image_strength "is just an alternate way to set step_schedule_start, which is done via the calculation 1 - image_strength." Its worked example is that image_strength=0.35 and step_schedule_start=0.65 are the same request.

The newer API repeats the split under different names. On /v2beta/stable-image/generate/sd3 the parameter is strength: "A value of 0 would yield an image that is identical to the input. A value of 1 would be as if you passed in no image at all." Four endpoints away, on /v2beta/stable-image/control/structure, it is control_strength: "How much influence, or control, the image has on the generation. Represented as a float between 0 and 1, where 0 is the least influence and 1 is the maximum."

Both are floats, both run 0 to 1, and both govern how much the input matters. They point in opposite directions, in one spec, under names differing by a word.

What is the parameter called in each tool?

The column that matters is the one on the right.

Twelve parameter names for one idea, three data types, two polarities. Three rows need a note.

Midjourney has no denoise parameter. It has --iw, which the Parameter List defines in six words: "Control the impact of image prompts". The Image Prompts article gives the direction: "By choosing a higher weight, your Image Prompt will have a bigger impact on the finished image." Its table publishes a default of 1, a range of 0 to 3 for Version 8.1 and Version 7, and 0 to 2 for Niji 7. Note the gap: the current default model is V8.2, the Feature Compatibility chart ticks Image Weight in the merged V8.1 and V8.2 column, and yet that range table has no V8.2 column. Our guide to Midjourney prompt weights covers the rest of the weighting vocabulary.

Ideogram counts in integers. image_weight runs 1 to 100: "How strongly the output should resemble the input image. Higher values keep the input image's structure; lower values allow the prompt to drive the output more." Version 3.0 defaults to 50; on 4.0 "the weight is chosen automatically from the edit instruction; supplying a value overrides that automatic choice."

Leonardo uses words. Its image-to-image guidance takes strength as LOW, MID or HIGH, defaulting to MID. It does not state which end preserves the input, so we will not state it either.

Which tools document no denoising strength at all?

Three of the most-used image models, which is worth knowing before you hunt for a slider that does not exist.

Google's Gemini image generation documentation contains zero occurrences of "strength" or "denoise". Its answer is a prompting technique it calls high-fidelity detail preservation: "To ensure critical details (like a face or logo) are preserved during an edit, describe them in great detail along with your edit request." The control lives in the sentence.

OpenAI's Images API exposes input_fidelity, documented as "Control how much effort the model will exert to match the style and features, especially facial features, of input images." It takes high and low, defaults to low, and its polarity is the inverse of strength.

Black Forest Labs publishes image_prompt_strength on the FLUX1.1 Ultra endpoint, "Strength of the image prompt influence on generation (0.0 to 1.0)", default 0.1. The FLUX.2 and FLUX Kontext request bodies have no strength field of any kind, and the word "denoising" appears nowhere in BFL's published API documentation.

Midjourney's Edit Model and Editor belong here too: instruction-based editing, inpainting and outpainting with no numeric control, plus a two-position Remix choice.

What do the bands actually mean?

Everything in this section is our own observation, not vendor documentation. No vendor we checked publishes a quality-versus-strength curve or a recommended-value table for general image-to-image work.

Described on the "higher means more change" scale, since most tools use it:

• 0.05 to 0.15, retouch. The picture comes back recognisably itself. Grain, blemishes and mild colour shifts move. Composition, faces and text do not.
• 0.15 to 0.35, structure-preserving restyle. Surfaces change, geometry holds. Material swaps, lighting adjustments and medium changes land here. Faces start to drift near the top.
• 0.35 to 0.60, reinterpretation with the composition intact. Layout survives, most else does not. The working range for sketch-to-render.
• 0.60 to 0.80, loose reinterpretation. Subject and rough mass survive; details and identity do not. Use it when the input is a mood reference, not a thing you are editing.
• 0.80 to 1.00, effectively text-to-image. Above roughly 0.9 you are paying for an upload that barely influences anything.

On the inverted scales, flip it. High --iw, high image_weight, high control_strength, high image_prompt_strength and input_fidelity: high all sit where the low numbers sit above.

Those boundaries are ours: where we stopped getting one behaviour and started getting the next, on our inputs, on the models we happened to be using. Treat them as a bracket for a sweep, not as settings.

The one genuinely documented band runs against all of it. Stability publishes, inside the strength description on generate/sd3: "For SD 3.5 Flash, the best results for image-to-image generation are achieved with a strength between .94 - .97." That is a vendor saying the right answer for one model is a value that would be pure noise on most others.

A worked ladder: one photograph, six settings

Same input throughout: a sharp, well-lit photograph of a ceramic mug on a wooden table. Same prompt. Only the number moves.

That --iw column is a rough alignment for orientation, not a conversion. Midjourney publishes no mapping to any other tool's scale.

Copy-paste prompts for the six common image-to-image jobs

Prompt and setting are one instruction, which is the part most guides split apart: at 0.2 you must describe what stays, and at 0.7 you must describe what appears. The same sentence will not do both.

Photo restyle

Restyle this photograph as a 1970s editorial magazine shot. Keep the subject,
pose, framing and background layout exactly as they are. Change only the colour
grade, film grain and lighting quality to warm tungsten with visible halation.
Do not move, resize or re-pose anything.
[start at 0.22 change-scale / --iw 2.3 / image_weight 78 / control_strength 0.78]

Convert this interior photo to a hand-painted gouache illustration. Preserve the
room's geometry, furniture placement and window position precisely. Replace only
surface rendering with visible brush texture and slightly flattened perspective.
[start at 0.35 change-scale / --iw 2.0 / image_weight 65]

Reinterpret this street photograph as a charcoal life drawing on toned paper.
Keep the composition and the position of every figure. Let all texture, colour
and material description be replaced by charcoal and chalk.
[start at 0.5 change-scale / --iw 1.5 / image_weight 50]

Turn this daylight exterior photograph into the same scene at night. Keep every
building, vehicle and person exactly where they are, at the same scale. Change
only the light sources, sky, reflections and the colour temperature.
[start at 0.3 change-scale / --iw 2.1]

Sketch to render

Render this pencil sketch as a photographic product shot. Follow the sketch's
composition, proportions and silhouette exactly. Add realistic material, studio
three-point lighting, contact shadow and a neutral seamless backdrop. Ignore all
construction lines, hatching and stray marks in the sketch.
[start at 0.55 change-scale / control_strength 0.45 on a structure endpoint]

Turn this architectural line drawing into a photoreal exterior render at golden
hour. Keep every wall position, window opening and roof line from the drawing.
Choose materials consistent with the labels written on the drawing.
[start at 0.6 change-scale / control_strength 0.4]

Render this rough wireframe as a finished UI screenshot. Keep the exact block
positions, proportions and reading order. Replace placeholder boxes with real
components, real type and a light neutral theme.
[start at 0.65 change-scale]

Convert this rough marker concept sketch into a clean vector-style illustration.
Preserve the silhouette and internal divisions exactly. Replace the marker
texture with flat fills, consistent stroke weight and no visible sketch lines.
[start at 0.45 change-scale / image_weight 55]

Colour change while holding structure

Change only the colour of the jacket in this photograph to deep forest green.
Every other element, including the fabric folds, stitching, lighting, shadows,
background and the subject's face, must remain identical in appearance.
Do not restyle, do not resample, do not re-light.
[start at 0.12 change-scale / --iw 2.8 / image_weight 88 / input_fidelity: high]

Recolour this packaging design from blue to terracotta. Preserve the logo shape,
all typography, kerning, layout and the die-line exactly. Only hue changes.
[start at 0.15 change-scale / image_weight 85]

Shift the overall colour temperature of this photograph from cool daylight to
warm late afternoon. Keep subject, composition, focus and grain untouched.
Adjust white balance and shadow tint only.
[start at 0.1 change-scale]

Change the wall paint in this room from off-white to deep clay. Keep every
furniture position, texture, shadow and reflection exactly as photographed, and
keep the light falloff across the wall consistent with the existing window.
[start at 0.18 change-scale / --iw 2.6]

Background swap

Replace the background behind the subject with a plain studio backdrop in warm
grey, lit from the upper left. Keep the subject's outline, pose, hair detail,
clothing and skin tone exactly as they are. Match the new background's light
direction to the existing light on the subject.
[start at 0.35 change-scale, or use a mask and skip strength entirely]

Keep this product exactly as photographed and place it on a sunlit marble
kitchen counter with a soft window reflection behind. The product's shape,
label, colour and contact shadow must not change.
[start at 0.3 change-scale / --iw 2.2]

Extend this photograph's background outward to a 16:9 frame. Continue the
existing wall, floor line and lighting gradient plausibly. Do not alter, move or
rescale anything already visible in the original frame.
[Stability outpaint, creativity 0.35]

For anything where the subject must survive intact, a mask beats a strength value. Our background replacement prompting guide has the per-tool syntax for the masked route.

Upscale and refine

Upscale this image and add plausible fine detail consistent with what is already
there: fabric weave, skin texture, surface grain. Invent nothing new. Do not
change composition, colour, expression or any element's position.
[start at 0.15 change-scale / Stability creativity 0.25]

Refine this low-resolution render at 2x. Sharpen edges and resolve small
geometry that is currently mushy. Preserve the exact camera position, materials
and lighting. Do not add objects.
[start at 0.2 change-scale / Stability creativity 0.3]

Second-pass refine this generated image. Fix soft focus and mushy detail in the
midground only. Keep the foreground subject and the overall composition
identical.
[hires-fix denoising strength 0.3, not the 0.7 default]

Clean up this scanned document photograph. Straighten the page, remove shadow
gradient and paper texture noise, and sharpen the text. Every character, line
break and mark on the page must remain exactly as it is.
[start at 0.1 change-scale / input_fidelity: high]

Character retouch

Retouch this portrait. Even out skin tone, reduce shine on the forehead and
remove flyaway hairs. The face's structure, proportions, expression, eye colour
and identity must be preserved with no alteration whatsoever.
[start at 0.08 change-scale / input_fidelity: high / image_weight 92]

Change this character's outfit to a charcoal wool overcoat. Keep the face, hair,
pose, body proportions and background exactly as they are. Only clothing changes.
[start at 0.25 change-scale / --iw 2.4]

Age this character forward by roughly ten years while keeping them recognisably
the same person. Preserve bone structure, eye shape, hairline position and the
existing lighting. Change skin texture and hair colour only.
[start at 0.3 change-scale, then sweep down before you sweep up]

Place this character in a different setting: a rain-wet street at night. Keep
the face, hair, clothing and body proportions unchanged, and relight the subject
to match the new scene without altering any feature.
[start at 0.4 change-scale / --iw 2.0 / image_weight 70]

Calibration prompts, for a model you have not used before

Same input, same prompt, four runs at 0.15 / 0.35 / 0.55 / 0.75 change-scale,
same seed. Write down which run first stopped looking like the input. That
number is this model's ceiling for edits, not the one from a tutorial.

Describe this image in enough detail that the description alone would reproduce
its composition, lighting and subject. I will use that description as the prompt
for an image-to-image run, so it must not contain instructions, only description.

Rewrite this edit instruction twice: once as a preservation instruction listing
everything that must not change, and once as a generation instruction describing
the finished image from scratch. I will pair the first with a low strength value
and the second with a high one.

Why does a value that works in one tool fail in another?

Three reasons, and only one is the polarity problem.

The step count moves underneath you. In diffusers, the steps actually run are your requested steps multiplied by the strength, so 20 steps at 0.5 gives you 10. Stable Diffusion WebUI does the same by default, and ships a setting to turn it off labelled "With img2img, do exactly the amount of steps the slider specifies", annotated "normally you'd do less with less denoising". ComfyUI does the opposite: it computes a longer schedule and runs your full step count from part-way along it, so a low denoise costs the same compute as a high one. Two front ends, one model, the same 0.4, a two-to-one difference in how much sampling happens. For what steps actually buy you, see what sampling iterations do.

Models disagree inside one vendor. Stability publishes its 0.94 to 0.97 recommendation for SD 3.5 Flash alone, on an endpoint whose other models take the same field with no recommendation at all. Ideogram's 3.0 and 4.0 remix endpoints default differently. Midjourney's --iw ceiling is 3 on V8.1 and 2 on Niji 7. A number is a property of a model, not of a concept.

The input image is half the variable, and none of it is documented. A flat, low-contrast input has less structure to lose, so a mid value wanders further than the same value on a crisp photograph. Resolution and aspect-ratio mismatch both change things before sampling starts. Change the kind of input you feed and your calibration is stale even though the model has not moved. When results break in ways a slider will not fix, our image prompt troubleshooting guide works through twenty failure modes.

What do the vendors not publish?

• Any quality-versus-strength curve, for any model, from any vendor we checked.
• Any recommended-value table for general image-to-image work. The SD 3.5 Flash note is the only documented band, and it covers one model.
• Midjourney's --iw range on V8.2, the current default, despite the compatibility chart marking image weight supported there.
• Which direction Leonardo's LOW to HIGH scale runs.
• Any mapping between one tool's scale and another's. Nobody publishes a conversion, us included.
• How the input image's own properties change the effective strength.
• Why ComfyUI's documentation table and its node tooltip state opposite polarity.

Where this page says a thing is not documented, that is what we found, not a gap in the checking.

How can you check this yourself?

Midjourney's docs.midjourney.com returns 403 to an ordinary fetch, which is why write-ups quote each other. Its Help Center JSON API answers a plain request, and the updated_at field is how you date a claim:

https://docs.midjourney.com/api/v2/help_center/en-us/articles/32040250122381.json

One trap cost us time: the marks in that Feature Compatibility chart are img elements, not characters, so stripping the tags leaves a table that looks empty. Read its raw HTML.

Stability's whole spec, including every description quoted above, is one document:

https://api.stability.ai/v2alpha/openapi

For diffusers and ComfyUI, read the source. A docstring and a tooltip are more current than any tutorial, and in ComfyUI's case more accurate than its own documentation table.

Worth saying plainly, because it is the boundary of what we do: Prompt Architects does not generate images and has no denoising strength slider. We generate the prompt. What breaks in this workflow is bookkeeping. Six variants of one instruction, each paired with a different number, across three tools whose numbers run in opposite directions, is the thing that lives in a dozen browser tabs until you lose the one that worked. Save the instruction once, make the subject and the setting variables, and the sweep becomes repeatable instead of remembered. Image prompt generation starts on the Pro plan at $4.99 a month at the time of writing, and the Free plan runs 5 prompt enhancements per day, forever, per our FAQ.

Sources and dates

Every parameter name, range, default and direction above was read from the vendor's own documentation or reference implementation on August 29, 2026.

Source code: diffusers img2img pipeline · ComfyUI nodes.py and comfy/samplers.py · Stable Diffusion WebUI ui.py, sd_samplers_common.py and shared_options.py.

Vendor docs: ComfyUI's KSampler page · Stability's API reference, whose client-rendered page loads the spec quoted above · docs.bfl.ai · Ideogram Remix · the openai/openai-openapi schema · Gemini image generation · Leonardo Create a Generation.

Midjourney, through the Zendesk JSON API because the HTML paths return 403: Image Prompts (updated August 28, 2026), Version (July 24, 2026, the compatibility chart and the V8.2 default), Parameter List (June 11, 2026), Edit Model (August 28, 2026) and Remix (July 27, 2026).

Adobe Firefly was omitted rather than guessed at: its published specification URL returned an application shell rather than the document, and quoting a summary of a page you could not read is how wrong parameters get published.

Product facts checked against our own live pages on August 29, 2026: /faq for the free daily allowance and for image prompt generation starting at Pro, and /pricing for the Pro price and the plan comparison.