self-eval-t2i

name	self-eval-t2i
title	Self-Evaluation Unlocks Any-Step T2I Generation
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.22374
keywords	["text-to-image","diffusion","training","few-step","self-supervised"]
description	Train text-to-image models from scratch for any inference step count via self-evaluation mechanism. Model evaluates its own generated samples using current score estimates as dynamic self-teacher, enabling global distribution matching without external teachers—achieving few-step quality equivalent to many-step models at all budgets.

Overview

Self-E addresses a fundamental limitation in text-to-image generation: achieving high quality at any inference budget requires either many steps or external teacher models. This framework enables from-scratch training where models teach themselves through self-evaluation, eliminating teacher dependence.

Core Technique

The key insight is that models can evaluate their own samples by leveraging internal score estimation during training.

Self-Evaluation Mechanism: The model generates samples and evaluates them using its current learned score function.

# Self-evaluating text-to-image generation
class SelfEvaluatingGenerator:
    def __init__(self, flow_model):
        self.model = flow_model
        self.optimizer = torch.optim.Adam(self.model.parameters())

    def training_step_with_self_evaluation(self, x_t, x_0, prompt):
        """
        Two complementary training objectives:
        1. Flow matching: local supervision via data
        2. Self-evaluation: global supervision from own scores
        """
        # Objective 1: Flow matching (standard)
        # Teaches model how samples should move toward data
        flow_loss = self.model.flow_matching_loss(x_t, x_0, prompt)

        # Objective 2: Self-evaluation (novel)
        # Use model's own score estimates for additional supervision
        self.eval()  # Freeze for score computation
        with torch.no_grad():
            # Generate samples during training
            x_gen_1 = self.model.sample(x_T=torch.randn_like(x_0), t=1.0)
            x_gen_2 = self.model.sample(x_T=torch.randn_like(x_0), t=0.5)

            # Compute score estimates for generated samples
            
            score_1 = .model.score_estimate(x_gen_1, prompt)
            score_2 = .model.score_estimate(x_gen_2, prompt)

        .train()  

        
        
        self_eval_loss = .classifier_free_guidance_loss(
            x_gen_1, x_gen_2, score_1, score_2, prompt
        )

        
        total_loss = flow_loss +  * self_eval_loss

        total_loss.backward()
        .optimizer.step()

         {: flow_loss, : self_eval_loss}