SOURCE · PROMPT · DENOISING · GEOMETRY · TEST

Transform a source image with img2img in Forge

Use the source as structure, the prompt as direction and denoising strength as a controlled departure—not as a magic quality setting.

00 · Direct answer

A controlled img2img pass has five contracts

The source does not work alone, and neither does the prompt.

01 · SOURCEChoose what may survive

Composition, silhouette, color and identity enter as encoded image details.

02 · DIRECTIONDescribe the new frame

The prompt names the intended whole-image result and stops reinforcing unwanted content.

03 · DEPARTUREProbe denoising

Separated values reveal where preservation becomes transformation.

04 · GEOMETRYResolve aspect ratio first

Target size and resize mode decide what image reaches diffusion.

05 · TESTCompare one variable

A fixed seed and checklist show which change produced the selected result.

Page boundary

This page owns whole-frame transformation. Use inpainting when a mask must protect or change a local area, ControlNet when pose or edges must become explicit constraints, outpainting when the canvas must expand, and upscaling when only pixel dimensions should grow.

01 · Transformation planner

Name what the source is allowed to control

A variation, restyle, subject change and iterative refinement require different preservation thresholds.

PRESERVE CONTRACT

Keep the source composition recognizable while allowing a new rendering of the whole frame.

Source role
Composition, subject placement, broad color and silhouette remain useful constraints.
Prompt role
Describe the result you want, including any attributes that should replace the source interpretation.
First probe
Start with separated low and middle denoising probes. Keep the source, seed and target dimensions fixed.
Done when
The new image is meaningfully different, but the source arrangement you intended to preserve is still attributable.

These are practical test routes synthesized from current code and earlier demonstrations. They are hypotheses to verify with the exact model package.

02 · Signal model

Four inputs negotiate every whole-frame result

When several inputs move together, the output cannot explain which one mattered.

SOURCE

Initial structure

Encoded pixels establish layout, color, shape and local details.

PROMPT

Target meaning

Text says what the transformed image should become.

SEED
4821

Noise path

A fixed seed makes comparisons attributable; a new seed explores another path.

DENOISING

Schedule segment

The selected sampling interval controls available departure from the source latent.

RESULTOne new whole-frame image with a measurable preservation/change trade-off
03 · Clean baseline

Build one attributable img2img transformation

Leave ControlNet, LoRAs, scripts and postprocessing out until this pass works.

  1. 01

    Prove the source and model separately

    Open or generate one normal image with the intended checkpoint, VAE and model family. Save the untouched source before any transformation.

  2. 02

    Open Img2img → img2img

    Upload the source or use the gallery button that sends the selected image and generation parameters to img2img. Confirm the expected gallery item reached the canvas.

  3. 03

    Set the target geometry first

    Use Resize to or Resize by. Match the source aspect ratio for the first baseline, then choose a resize mode deliberately if the ratio must change.

  4. 04

    Describe the new whole-frame result

    Write what the finished image should show. Keep source details that must survive and remove prompt terms that fight the requested transformation.

  5. 05

    Freeze a reproducible comparison

    Record model, VAE, prompt, negative prompt, seed, sampler, scheduler, steps, CFG, source dimensions, output dimensions and resize mode.

  6. 06

    Run separated denoising probes

    Compare 0.25, 0.50 and 0.75 as a practical diagnostic ladder—not universal quality settings. Change no other control.

  7. 07

    Select the smallest sufficient departure

    Keep the lowest probe that completes the task. If the old and new subjects mix, test a higher interval; if structure disappears, test lower.

  8. 08

    Verify and save the pass boundary

    Compare source and result for composition, subject, identity, edges, color and output dimensions. Save the exact result used for the next iteration.

DONE STATEThe requested whole-frame change is visible, the intended source structure remains attributable, output geometry is correct, and the recorded setup reproduces the selected pass.
04 · Denoising map

Find the preservation threshold with separated probes

The visual meters are a diagnostic model, not measured similarity or a quality score.

DENOISING PROBE0.25
Source retention hypothesis88%
Prompt freedom hypothesis28%

A light probe for cleanup, restrained restyling or finding whether the source already contains most of the desired answer.

Code boundary: current Forge accepts 0–1. Sampling uses denoising to choose the img2img schedule segment. The optional fix-steps setting changes how displayed steps map to that segment.

05 · Resize geometry

Choose what happens when aspect ratios disagree

Resize mode changes the input before diffusion; denoising cannot undo a geometry choice reliably.

Stretch to target

Just resize

Resizes the source directly to the requested width and height. A changed aspect ratio compresses or stretches shapes.

Fill, then center-crop

Crop and resize

Preserves aspect ratio, fills the target frame and removes centered overflow. Important content near the source edge can be lost.

Fit, then fill bands

Resize and fill

Preserves aspect ratio, fits the source inside the target and fills the remaining bands using stretched edge data before diffusion.

Encode, then resize latent

Just resize (latent upscale)

Skips the ordinary pixel resize, encodes the source, then bilinearly interpolates its latent to the target size. It is not a selectable ESRGAN-style pixel upscaler.

RESIZE TOExplicit width × height64–2048 · step 8
RESIZE BYSource × scale0.05–4.00 · step 0.01
06 · Technical nuances

Know what each current control changes

Labels, ranges and processing behavior are pinned to the current Forge commit.

Source image

Ordinary img2img transforms the whole source. The current processing path flattens RGBA input onto the configured img2img background color, resizes it unless latent upscale is selected, and encodes it into the initial latent.

Denoising strength

The current UI accepts 0–1 in 0.01 steps and displays 0.75 by default. Sampling code uses it to choose the img2img schedule segment; it is not an opacity slider or a universal percentage of similarity.

Sampling steps

By default, lower denoising generally uses a shorter portion of the selected step schedule. The optional setting “With img2img, do exactly the amount of steps the slider specifies” changes that calculation. Record this setting when comparing environments.

Prompt and negative prompt

The source supplies image structure; text conditioning supplies the requested interpretation. A prompt that still names the old subject can resist a subject replacement even when denoising is raised.

Seed

The seed controls the noise path added to the encoded source. Holding it fixed makes denoising comparisons attributable; changing it explores another path but can hide the effect of the control you meant to test.

Resize to

Width and Height each expose 64–2048 in steps of 8, with displayed values of 512. Auto detect size is a button; automatic updates after loading are a separate setting and are off in current defaults.

Resize by

Scale exposes 0.05–4.00 in steps of 0.01 and previews source-to-output resolution. For a single image, Forge derives width and height from the source and rounds them down to multiples of 8.

Resize mode

Current choices are Just resize, Crop and resize, Resize and fill, and Just resize (latent upscale). They decide what reaches diffusion before denoising; they do not merely change the saved canvas.

Batch count and size

Batch count creates repeated iterations. Batch size displays 1–8. More outputs do not make one comparison more controlled; keep a fixed seed for diagnosis and use batches only after one image works.

CFG and model-specific controls

Current Img2img exposes CFG Scale 1–30, Distilled CFG Scale 0–30 and a normally hidden Image CFG Scale. Useful values depend on the loaded model contract; this page does not promote one cross-family default.

For a single image, Resize by calculates width = source width × scale and height = source height × scale, then rounds both down to a multiple of 8 before processing.

Inspect the current request path ↗
07 · Reproduction checklist

Save the pass before you iterate from it

Sending a selected result back to img2img creates a new source boundary.

Img2img reproduction checklist
RECORDED0 / 11 recorded
SOURCE AUntouched inputsaved outside the canvas
PASS BOne controlled changefixed seed · measured geometry
SOURCE BSelected outputnew checklist before the next pass
08 · Failure router

Fix the first contract that stopped matching

Similarity, geometry and output failures belong to different diagnostic branches.

The output is almost identical to the source

Confirm that the prompt actually describes a change, then raise denoising in one controlled step. Check whether an imported seed, model or extension state differs from the comparison you intended.

The source composition disappears

Lower denoising, restore the fixed seed and remove unrelated prompt or extension changes. If exact pose or edges are mandatory, prove the requirement with ControlNet instead of asking ordinary img2img to preserve everything.

Continue with the focused guide →
The image is stretched or squashed

Compare source and target aspect ratios. Just resize forces the source into the target geometry. Match the ratio, or choose Crop and resize / Resize and fill according to whether losing edges or adding bands is acceptable.

Inspect the details ↗
Important content is cut off

Crop and resize center-crops overflow after filling the target. Switch to a matching target ratio, reposition the source before Forge, or test Resize and fill when preserving all source edges matters.

Inspect the details ↗
Filled bands look repeated or smeared

Resize and fill constructs the missing bands from stretched edge data before diffusion. Use a closer aspect ratio, raise denoising only if the prompt can plausibly rebuild the bands, or move to an outpainting workflow for deliberate canvas extension.

Continue with the focused guide →
The new subject looks like a hybrid of old and new

Remove the old subject from the prompt, keep the seed fixed and test a higher denoising interval. An earlier FLUX demonstration shows the same transitional failure; its numeric values are examples, not a model-independent recipe.

Inspect the details ↗
The result changes unpredictably after Send to img2img

The current button explicitly sends the image and generation parameters. Verify prompt, seed, dimensions, sampler, steps and extension state before generating; do not assume only pixels crossed the tab boundary.

Inspect the details ↗
Img2img makes a different image when I only wanted more pixels

Img2img is a diffusion regeneration path. Resize settings can invoke a configured pixel upscaler before diffusion, but denoising still permits a new image. Use Extras when the dominant task is pixel enlargement without a new diffusion interpretation.

Inspect the details ↗
The result is black, blank or corrupted

Stop adjusting denoising. Prove the base checkpoint and VAE, remove extensions and route the first bad stage through the dedicated output diagnostic.

Continue with the focused guide →
09 · Advanced boundaries

Add batches and hidden settings after one image works

These controls change input handling or sampling semantics and belong in the checklist.

BATCH INPUT

Upload or server directory

Current Batch accepts uploaded files or a directory on the Forge host. Directory mode is disabled by --hide-ui-dir-config. Each input receives the configured iterations and batch size.

Inspect batch processing ↗
PNG PARAMETERS

Append selected PNG info

Batch can append source prompts and read Seed, CFG, Sampler, Steps and Model hash. “Append” matters: the UI prompt remains and selected source text is added to it.

Inspect PNG-info handling ↗
SETTINGS

Noise, color and exact steps

Settings expose initial noise multiplier, extra noise, color correction, exact img2img steps, transparent-input background color, autosize and an img2img upscaler. Defaults are not a universal recipe.

Inspect current options ↗
ROUTE CONFUSION

Resize is not “just upscale”

Img2img regenerates an image from the initial pixels and prompt. Use a dedicated upscale route when you want more pixels without a new diffusion interpretation.

Open the related issue ↗
10 · Img2img FAQ

Questions between the source and the next pass

Answers separate current code behavior, earlier demonstrations and model-specific unknowns.

Where is img2img in Stable Diffusion WebUI Forge?

Open Img2img, then choose the img2img sub-tab. Upload a source image or use the gallery button that sends the selected image and its generation parameters to img2img.

What does img2img do in Forge?

It encodes a source image, adds a seed-dependent noise path according to denoising strength, and samples a new result under the prompt and model conditioning. Ordinary img2img applies this transformation to the whole frame.

What is the difference between txt2img and img2img?

Txt2img starts without source pixels. Img2img starts from an encoded source image, so its composition, colors and shapes can influence the result depending on denoising and geometry.

What is the difference between img2img and inpainting?

Ordinary img2img transforms the whole source. Inpainting adds a mask so a region or its inverse becomes the edit contract. Use the separate inpainting guide for local repair, removal or replacement.

What controls the result: the source image or the prompt?

Both. The source supplies initial spatial and visual details; the prompt supplies the requested interpretation. Denoising strength controls how much sampling freedom exists to move away from the encoded source.

What does denoising strength do in Forge img2img?

It selects how much of the img2img noise and sampling schedule is used. Lower values usually preserve more source structure; higher values permit larger change, but the relationship is not a literal similarity percentage.

What happens at denoising strength 0?

The current code accepts 0, but the sampling path has effectively no denoising segment. Treat it as a diagnostic boundary for the encoded/resized source—not a useful transformation setting or a guarantee of pixel-identical output.

What happens at denoising strength 1?

The UI accepts 1, while sampling code caps its schedule calculation just below 1. This gives very high departure from the source; composition or identity can disappear.

What denoising strength should I use for img2img?

There is no universal value. Compare separated probes with the same source, prompt, seed, model, sampler and dimensions, then narrow the interval that completes the task with the least unnecessary drift.

Why does img2img return almost the same image?

The prompt may not request a visible change, denoising may be too low for that task, or the fixed source signal may dominate. Raise only denoising first and compare against the same checklist.

Why does img2img change too much?

Lower denoising and restore a fixed seed. Confirm that target dimensions, resize mode, prompt, model and extensions did not also change when the image was sent or uploaded.

Will the same seed reproduce the same img2img image?

Only when the source and the rest of the effective setup also match: model files, prompt, dimensions, resize mode, sampler, scheduler, steps, denoising, CFG, precision and enabled add-ons.

What does the seed change in img2img?

It changes the noise path applied to the encoded source. A different seed can produce another valid interpretation at the same denoising strength, so keep it fixed while diagnosing denoising.

How do I preserve a face in img2img?

Use lower denoising, matching geometry and a prompt that does not contradict identity. Ordinary img2img cannot guarantee identity; use inpainting for a protected local region or a separately confirmed identity tool when exact likeness is required.

How do I preserve pose and composition?

Match the source aspect ratio, start with lower denoising and keep the seed fixed. If pose or edges must be explicit constraints rather than tendencies, add one proven ControlNet unit after the base img2img pass works.

How do I change only the style of an image?

Keep the source composition, replace contradictory style terms in the prompt and compare low-to-middle denoising probes. Verify that the target treatment is consistent across the whole frame.

How do I change one subject into another?

Name the new subject, remove the old one from the prompt and compare wider denoising probes. Expect a transition interval where old and new visual traits mix before replacement becomes clear.

Which img2img resize mode should I use?

Match the source ratio whenever possible. Otherwise use Just resize only when distortion is acceptable, Crop and resize when losing centered overflow is acceptable, or Resize and fill when preserving the full source is more important than clean new bands.

What does Just resize do?

It resizes the source directly to the target width and height. If source and target aspect ratios differ, shapes are stretched or compressed before diffusion.

What does Crop and resize do?

It preserves aspect ratio, enlarges the source until the target is filled, centers it, and crops the overflow. Edge content may be removed.

What does Resize and fill do?

It preserves aspect ratio, fits the source inside the target, and fills empty bands from stretched edge data before diffusion processes the image. It is not the same as intentional outpainting.

What does Just resize (latent upscale) do?

Forge skips the ordinary pixel resize, encodes the source, then bilinearly interpolates the latent to the target dimensions. It does not expose a choice of ESRGAN-style upscaler in that mode.

What is the difference between Resize to and Resize by?

Resize to sets an explicit target width and height. Resize by derives the target from the source dimensions and a 0.05–4.00 scale value; the single-image path rounds dimensions down to multiples of 8.

How do I copy the source image dimensions?

Use the Auto detect size from img2img button. Automatic width and height updates on image load are controlled by a separate setting that is off in the current defaults.

Can img2img upscale an image?

It can output larger dimensions and may run the configured img2img pixel upscaler during ordinary resize, but diffusion still regenerates the image. Use the upcoming upscale guide or Extras when enlargement—not reinterpretation—is the dominant task.

Can Forge batch img2img images?

Yes. The Batch sub-tab accepts uploaded files or a server directory when directory access is allowed. Prove one image first, then record batch count, batch size, source handling and output naming.

What does Append png info to prompts do in batch img2img?

For selected properties, Forge reads PNG information and appends source Prompt and Negative prompt to the UI prompts; it can also take Seed, CFG scale, Sampler, Steps and Model hash. Review the resulting prompt before a large run.

Does img2img work with FLUX in Forge?

An earlier video demonstrates one FLUX Dev NF4 setup. That proves one historical combination, not arbitrary current FLUX, GGUF, NF4, LoRA or extension compatibility. Record the complete model package and treat untested combinations as VARIES BY SETUP.

Does img2img require a special checkpoint?

No special img2img filename is required for the core route, but the checkpoint, VAE, text encoders, preset and add-ons must form a compatible model package. Follow the model publisher for family-specific requirements.

How do I make an img2img result reproducible?

Save the source and selected output; record Forge commit, model files, prompts, seed, sampler, scheduler, steps, CFG, source and target dimensions, resize mode, denoising, fix-steps state, extensions and every result chosen for another pass.