Multi-turn image edits stop smearing: TailSteer steers only the denoising tail, cutting cumulative drift over 15 to 20 turns
Synopsis
In multi-turn image editing, each output becomes the next input, so small unintended changes to untouched content accumulate into visible drift. TailSteer concentrates preservation corrections near the tail of the diffusion denoising trajectory while keeping early dynamics close to the base model, and learns them from the editor's own on-policy rollouts using unchanged source content as reference, without curated editing pairs. Across four backbones in the FLUX, SenseNova and Qwen families, cumulative drift over 15 to 20 editing turns drops substantially while text-to-image and single-turn editing benchmarks stay largely intact.
Interpretation
Reliability in multi-turn image editing no longer has to be inferred from single-turn skill: a post-training step that touches only the end of sampling keeps subject and background detail stable after more than a dozen repeated edits. Earlier routes to multi-turn consistency added auxiliary modules such as masks or frequency corrections, reshaped the latent representation, cached visual memory across turns, or retrained on multi-turn data and cross-turn reinforcement learning; these either constrain the base model's generative prior or do not improve the model itself. On FLUX.2 [klein] Base 9B with 50 sampling steps per edit, running the base model for the first 49 steps and the terminally fine-tuned model only for the final update markedly reduced visible drift and kept results visually stable through 20 turns, where the base model showed facial degradation after five.
The full method spreads the correction across the tail of the denoising trajectory rather than piling it onto the last step, so one trained model runs the entire sampling process on its own. With last-step-only training, parameters are shared across timesteps, so terminal supervision disturbs earlier learned dynamics; applying that model to all 50 steps produced outputs with fewer conspicuous artifacts than the base model but desaturated and overly smooth, which required explicit model switching. Training starts from the frozen base model, asks the refined copy to reproduce base velocities in the prior region, and from a chosen steering onset blends gradually toward a preservation endpoint defined by unchanged source regions, with supervision states taken from the refined model's own on-policy denoising rollouts; inference needs no base model, masks or extra adapters.
Perspective
The demonstrated setting is multi-turn editing with standard image feedback on four open-weight backbones (FLUX.1 Kontext [dev], FLUX.2 [klein] Base 9B, SenseNova-U1.5-8B-MoT and Qwen-Image-2.1), with training data drawn from Open Images and MONET source images paired with local-edit and identity-edit instructions; the intended reader works on image generation and editing systems.
Instruction-following gains are less consistent than preservation consistency, and multi-turn success rates improve only at selected steering onsets; the best preservation appears at earlier or intermediate onsets rather than the latest ones; VAE-based editors still show residual color shifts, which the authors attribute to repeated VAE reconstruction between turns and suggest latent feedback or better VAE fidelity could complement.