Zero to Director #6 — Talk to Your Image: Reference & Editing with Qwen-Image 2.1
From Last Time
Five posts in, every handle you own still passes through sentences. The reference bust of post #5 showed the carver a look — but "warmer light on the left, leave the heron alone" is still a translation problem: the translator reads words, the picture stays unsaid.
This post changes the machine itself. New carver, new caster — and the headline: a new translator that reads pictures. Same strokes though: a sampler, a dial, a seed. The workshop grammar you already speak carries over intact.
Why a Multimodal Blueprint
Editing has always been the awkward case for text-to-image. "Make the sign say OPEN" is not a description of a picture — it is an instruction about a picture, and until now the blueprint had no way to contain the picture it was talking about.
Qwen-Image 2.1 fixes the sentence, not just the vocabulary: the blueprint now holds up to sixteen image slots alongside the words.
<image1> is the image being edited; <image2> onward are references you point at mid-sentence.
The translator sees what you see, and the blueprint stops being a caption — it becomes a conversation with the photograph on the bench.
The New Stack
Three files, one machine, about thirty gigabytes in the int8 family:
- The carver —
qwen_image_2.1_int8_convrot.safetensors, the diffusion model itself. - The translator —
qwen3vl_8b_int8_convrot.safetensors, the picture-reading encoder that replaced the sentence-only one. - The caster —
qwen_image_2.1_vae_bf16.safetensors, its own decode step.
Two dial defaults change shape: steps 25 at full rebuild, and — the editing gift — 4 to 8 steps when the edit is local. Force stays at 1, the Turbo-family trait you know from post #2. The stations look familiar too: Load, Source, Condition, Sample, Decode, Save — the Condition station just got eyes.
The Language of Edits
The grammar that makes this machine sing:
- Instruct, don't describe. "Make the sunbeam warmer" beats "a workshop with warm sunlight" — you are talking to a picture that already exists, not commissioning a new one.
- Name what you can see. The translator reads the slots; "the metal sign" resolves because it can see the sign. Abstract nouns it cannot find in the image are guesses.
- Scope the edit. "Only the background," "leave the heron untouched" — the model honors boundaries stated in words better than boundaries you merely hoped for.
- Point at references mid-sentence. "Put the pot from
<image2>on the bench" uses a second slot as a noun, the way post #5's bust was a noun — but spoken inside the blueprint.
Examples
Retouch by words
06-retouch-by-words.json is the heron photograph on the new bench, edited in one sentence: light and atmosphere only, structure untouched.
Four to eight steps is the whole job: {OBSERVE-RETOUCH}
Cast from references
06-cast-from-refs.json points mid-sentence: the subject from <image1>, the scene from <image2> — multi-reference composition as one instruction.
The bust of post #5 lived outside the blueprint; this reference lives inside it: {OBSERVE-CAST}
Fix the sign
06-fix-the-sign.json is the party trick with a practical core — in-image text editing: the sign in the scene, re-lettered by instruction, pixels around it held still.
Writing legible text has been diffusion's oldest shame; this is the counterexample: {OBSERVE-SIGN}
Where the bust still wins
06-ref-vs-ipadapter.json runs one job both ways — post #5's IP-Adapter style route against this post's conversational route.
The honest boundary, measured: {OBSERVE-BOUNDARY}
Pitfalls
- Describing when you should instruct. Prompt habits from text-to-image leak in: "a bronze heron in a workshop" regenerates; "make this heron bronze" edits. The grammar is the tool.
- Unscoped edits drift. An instruction without a boundary is a suggestion — everything leans. Say what stays.
- Sixteen slots is a budget, not a suggestion. Every reference costs attention; a pile of slots muddies which noun binds to which image. Point, don't pile.
- Local edit, global habits. 25 steps on a one-word fix is a full recarve — pay 4 to 8 and keep the change small on purpose.
Your Move
Put a photograph you actually want changed into <image1> and say the change in one sentence — scoped, named, and instructed.
Then fix a sign, any sign: the first legible re-letter you do by conversation will spoil you for every eraser tool you have ever owned.
Next post goes back to the old machine with a smaller toy: boxes of chisels — LoRA — that specialize a carver you already trust.