Image-to-image (I2I) generation starts from a picture you supply and redraws it according to a prompt — keeping the composition and subject you liked while changing what you ask it to change. GLM-Image supports image-to-image alongside text-to-image, and together the two modes form one workflow: create from a prompt, then refine from a picture. This guide covers when I2I is the right tool, how to write prompts for it, the strength setting that governs everything, and the failure modes that waste generations.
The short version: use image-to-image whenever you already have an image that is 80% right. Re-rolling from a text prompt and hoping is how credits evaporate.
<!--IMG: The same product photo restyled through three image-to-image passes: studio, lifestyle, and seasonal-->
What is image-to-image generation?
Text-to-image turns words into a picture from nothing. Image-to-image turns a picture plus words into a new picture. You provide a reference image and an instruction; the model reads the reference, understands its structure, and regenerates it with your requested changes while holding onto what it can.
Mechanically, the reference image is encoded into the model's latent representation, your prompt steers how that representation gets rebuilt, and a strength setting controls how far the rebuild is allowed to drift from the original. Low drift means small touch-ups; high drift means reinvention.
What survives an I2I pass, in rough order of reliability: overall composition, the placement of major elements, the identity of the main subject (at low-to-medium strength), and the general color story. What changes: whatever you explicitly ask for, plus — this is the catch — whatever drift leaks in at higher strengths. I2I is a negotiation, not a mask. That is both its power and its risk.
When should you use image-to-image instead of text-to-image?
Four situations, in descending order of how often they apply:
Editing a near-miss. Your generation got the composition, the product, and the mood right, but the background is wrong or the lighting is flat. Regenerating from the prompt gambles away everything that worked; an I2I pass with "replace the background with X, keep the subject unchanged" preserves your win.
Style transfer on a solid composition. You have a photo or render whose layout is right, and you need it as a watercolor, an oil painting, or flat vector art. I2I re-skins the structure instead of inventing a new one.
Variations from a winning base. One hero image, six derivatives: seasonal versions, alternate colorways, different backgrounds for A/B testing. Each variation keeps the family resemblance because they all descend from the same reference.
Consistency across a series. A character that must look like herself across scenes, or a product that must stay identical across a catalog. Anchoring every generation to the same reference narrows the variance dramatically compared to prompt-only generation.
The general rule: the moment you find yourself re-rolling a text prompt more than two or three times to get back something you already had, you wanted image-to-image.
How to write an image-to-image prompt
I2I prompts follow a three-part pattern:
[what to change] + [what to keep] + [style and quality direction]
The middle part is the one people skip, and skipping it is the number-one cause of disappointment. The model will not automatically know which parts of the reference you consider sacred. Name them:
Replace the background with a Tokyo street at night, keep the product
and its label exactly unchanged, neon reflections on the surface,
commercial product photography, sharp focus
Convert to watercolor style, preserve the composition and all major
elements, soft washes, visible paper texture, museum-grade illustration
Change the lighting to golden hour, keep the model's pose and outfit
identical, warm rim light from the left, editorial photography
Notice what these do not do: they do not describe the image from scratch. The reference already carries that information. Your prompt's job is to state the delta — the change — and fence off what must not move. If you find yourself writing a full description of the picture, you are doing text-to-image with extra steps. The underlying craft is the same as prompt-writing in general — subject, style, concrete detail, covered in full in how to write AI image prompts — but the responsibility shifts: the reference supplies the scene, and the prompt supplies the edit.
Strength: the one setting that decides the outcome
The strength parameter (some tools call it denoising strength) controls how much of the reference the model is allowed to overwrite:
| Strength | Behavior | Best for |
|---|---|---|
| Low (roughly 0.3–0.4) | Light touch; identity and details survive | Retouching, lighting tweaks, texture |
| Medium (roughly 0.5–0.7) | Balanced edit; structure holds, content shifts | Background swaps, style shifts |
| High (roughly 0.7–0.9) | Reference becomes a suggestion | Reinvention, dramatic restyling |
Two honest notes. First, exact scales and parameter names differ between tools and versions — treat the ranges above as a starting grid and calibrate with your own tests. Second, strength interacts with resolution: the same setting can feel more destructive at small sizes, where fine detail has less room to survive.
The classic contradiction to avoid: high strength plus "keep everything the same." You are simultaneously telling the model to rewrite aggressively and to change nothing. Pick one. Identity preservation wants low strength; transformation wants high strength; most real edits live in between.
<!--IMG: The same reference image processed at low, medium, and high image-to-image strength-->
An iteration workflow that actually converges
Random one-shot editing produces random results. This loop converges:
- Create the base. Generate with text-to-image (or upload your own image). Spend your picks here — a strong base repays itself in every later pass.
- Diagnose before you edit. Write down what is wrong, specifically. "Lighting is flat" is an instruction. "Make it better" is a wish.
- First I2I pass — broad strokes. Background, style, or overall lighting at medium strength. One theme per pass.
- Second pass — specifics. Color grading, mood, detail emphasis at low strength.
- Feed the output back as the new reference. Each pass compounds: yesterday's result is tomorrow's starting point.
- Stop when the fixes get smaller than the drift. When a pass introduces as many new problems as it solves, you have converged — or the base needs rethinking, not another pass.
Two accelerators: reuse the seed when your tool exposes it, so a prompt tweak produces a controlled comparison rather than a new dice roll; and keep notes of which strength values produced which character of change, because your own calibration table is worth more than anyone's generic one.
Common failure modes and their fixes
The kitchen-sink prompt. "Change the background, make it night, restyle as painting, add snow, remove the car." Five edits at once means five chances to break something, and no way to know which instruction caused the damage. Fix: one theme per pass, in sequence.
Vague improvement requests. "Make it more professional" gives the model a mood with no lever. Fix: name the defect and the remedy — "replace the cluttered background with a clean gradient, keep the subject unchanged."
Fighting the reference. Asking for changes that contradict the composition ("make the subject fill the frame" on a wide landscape) forces the model to choose between your prompt and its anchor. Fix: when the layout itself is wrong, go back to text-to-image and fix the layout there.
Text edits at high strength. In-image text re-randomizes when the redraw is aggressive — the sign you liked comes back with new, wrong letterforms. Fix: redo text-bearing images at low strength, restate the exact text in the prompt, and keep it short. GLM-Image's text rendering is a strength here — it is the model's signature capability, and the deeper playbook is in best AI text rendering — but only when you work with it, not against it.
Infinite refinement. Pass ten rarely fixes what pass three couldn't. Fix: accept convergence, or restart from a better base. Finished beats perfect.
What image-to-image cannot do
I2I is redraw-with-instructions, not pixel editing. It cannot "move this 3 pixels left," recolor one exact pixel region on demand, or guarantee that your brand's hex codes and exact typography survive regeneration — that precision belongs to Photoshop, Illustrator, or Inkscape after the AI pass. Identity drift at high strength is physics, not a bug you can prompt away. And long in-image text stays risky on every model; short phrases are the reliable regime. The professional pattern is exactly this order: AI for the image, design tools for the final millimeters.
Frequently Asked Questions
Does GLM-Image support image-to-image?
Yes — both text-to-image and image-to-image are supported. The standard workflow is text-to-image for the base, then I2I passes for refinement and variations.
What strength should I start with?
Medium (roughly 0.5–0.7) for a typical edit. Drop to low when identity or detail preservation matters most; go high only for deliberate reinvention, and expect the reference to survive as a suggestion rather than a structure.
Can I use my own photos as references?
Generally yes — style transfers and edits on your own images are the core use case. For images of identifiable people, you are responsible for having the rights and consent to use them.
When is re-prompting better than another I2I pass?
When the flaw is in the composition itself. I2I inherits the reference's layout; it cannot fix a layout it is simultaneously preserving. Structural problems go back to text-to-image.
Try the loop on your own work: open the image generator →
