---
title: 'DreamPoster: Unified T2I Poster Generation'
url: https://www.emergentmind.com/topics/dreamposter
type: topic
---

# DreamPoster: Unified T2I Poster Generation

DreamPoster is a unified text-to-image (T2I) generation framework designed for synthesizing high-quality posters from user-provided source images and textual prompts, with explicit support for content fidelity, flexible resolution, and diverse layout outputs. Built on the Seedream 3.0 backbone, DreamPoster integrates systematic data annotation and a progressive, multi-stage training regimen. It achieves state-of-the-art usability and design quality benchmarks, significantly outperforming both GPT-4o and SeedEdit 3.0 in empirical evaluations [2507.04218].

## 1. Seedream 3.0 Diffusion-Transformer Backbone

DreamPoster adopts the Seedream 3.0 architecture, utilizing a diffusion–transformer (DiT) backbone that unifies three input modalities: a source image $I$, a prompt $T$ (typically title, tagline, etc.), and a noisy target latent $x_t$. Each modality is embedded as follows:

- **Image Embedding**: The input image $I$ is partitioned into $M$ patches. Each patch is projected into a $d$-dimensional space, forming $E_{\rm img} = [p_1, \dots, p_M] \in \mathbb{R}^{M \times d}$.
- **Text Embedding**: The textual prompt $T$ is processed using a standard encoder (e.g., RoBERTa/T5) to yield $E_{\rm txt} = [t_1, \dots, t_L] \in \mathbb{R}^{L \times d}$.
- **Noisy-Latent Embedding**: The current diffusion latent $x_t$ is flattened and projected to obtain $E_{\rm lat} = [\ell_1, \dots, \ell_N] \in \mathbb{R}^{N \times d}$.

Learnable positional embeddings $P_{\rm img}$, $P_{\rm txt}$, $P_{\rm lat}$ are added to each token according to modality. These are concatenated:
$$
H^0 = [E_{\rm img} + P_{\rm img};\, E_{\rm txt} + P_{\rm txt};\, E_{\rm lat} + P_{\rm lat}] \in \mathbb{R}^{(M+L+N)\times d}
$$
The sequence $H^0$ is passed through $L$ transformer diffusion blocks, each of which applies LayerNorm, MultiHeadSelfAttention, and a feedforward MLP to jointly attend over all modalities. The model predicts the denoising term
$\epsilon_\theta(x_t, E_{\rm img}, E_{\rm txt}, t)$,
optimized under a standard DDPM objective:
$$
\mathcal{L}_{\rm denoise} = \mathbb{E}_{t\sim [1,T],\,x_0,\,\epsilon \sim \mathcal{N}(0,I)} \left\| \epsilon - \epsilon_\theta(x_t, E_{\rm img}, E_{\rm txt}, t) \right\|^2.
$$

## 2. Data Annotation and Poster Curation Pipeline

The data annotation pipeline is engineered to yield well-structured triplets $(I_{\rm src}, T_{\rm src}, I_{\rm poster})$ and detailed typographic/layout metadata:

- **Raw-Poster Filtering**: Poster candidates undergo OCR (PP-OCR [Du et al. 2020]) and are given an aesthetic score $s_{\rm aes}(I)\in[0,1]$ from a pretrained network. Only posters with at least one OCR box and $s_{\rm aes}(I) \geq \tau$ are retained.
- **Source-Target Pairing**:
   1. Extract all text-box masks $\{m_i\}$.
   2. Inpaint the union of text regions $M = \bigcup_i m_i$ from $I_{\rm poster}$ to obtain a clean background $I_{\rm bg} = \mathrm{Inpaint}(I_{\rm poster}, M)$.
   3. Optionally segment the main subject as $I_{\rm subj}$.
   4. Concatenate all recognized OCR strings to yield $T_{\rm src}$.
   5. The triplet is: $I_{\rm src} = I_{\rm bg} \oplus I_{\rm subj},\  T_{\rm src},\  I_{\rm poster}$.
- **Poster-Captioner**: An image captioning model outputs:
   - Glyph-caption $C_{\rm glyph}$: font, size, color, and position for each string.
   - Layout-caption $C_{\rm layout}$: global layout properties (e.g., “title at top center”).
   Supervision is based on human-verified templates.

A high-level pseudocode is as follows:
```python
for I_poster in raw_dataset:
    boxes, strings = OCR(I_poster)
    if not boxes or aesthetic(I_poster) < τ: continue
    M = union(boxes)
    I_bg = Inpaint(I_poster, M)
    I_subj = SegmentMainSubject(I_poster)
    T_src = concat(strings)
    save_pair(I_bg, I_subj, T_src, I_poster)
    C_glyph, C_layout = PosterCaptioner.label(I_poster)
    save_annotations(C_glyph, C_layout)
```

## 3. Progressive Multi-Stage Training Regimen

DreamPoster employs a three-stage progressive curriculum:

1. **Stage 1 — Single-Task Text-Addition**: Trained only on pairs where the poster differs by overlaying $T_{\rm src}$ on $I_{\rm src}$, using $\mathcal L_{\rm denoise}$ conditioned on $(I_{\rm src}, T_{\rm src})$.
2. **Stage 2 — Multi-Task Mixed Training**: Incorporates five subtasks—addition, deletion, modification, restyling, layout shift. Each subtask is represented by a task tag $\tau_i \in \{1,2,3,4,5\}$ encoded as extra tokens. The loss is
   $$
   \mathcal{L}_{\rm stage2}
   = \sum_{i=1}^{5} \lambda_i\, \mathbb{E}_{\tau=\tau_i} \left\| \epsilon - \epsilon_\theta(x_t, E_{\rm img}, E_{\rm txt}, \tau_i, t) \right\|^2.
   $$
3. **Stage 3 — Fine-Grained Aesthetic Alignment**: Fine-tuning on a curated set of highly rated posters, the objective adds an aesthetic-alignment regularization. With $s_{\rm aes}(\cdot)$ a differentiable aesthetic network,
   $$
   \mathcal{L}_{\rm stage3}
     = \mathcal{L}_{\rm stage2}
     + \alpha [1 - s_{\rm aes}(G_\theta(I_{\rm src}, T_{\rm src}))]
   $$
   where $\alpha$ is a hyperparameter. Each stage fine-tunes from the previous weights, using Adam and learning rate decay.

## 4. Evaluation Methodology and Empirical Results

DreamPoster is evaluated on a test set of 500 held-out poster scenarios covering movie, product promo, holiday card, and event flyer applications. Evaluation utilizes both expert designer rating and large-scale user studies:

- **Expert Ratings** (40 designers):
  - Prompt Following (PF) ∈ [1,5]
  - Subject Preservation (SP) ∈ [1,5]
  - Design Sense (DS) ∈ [1,5]
- **User Study** (40 users, 30 designers):
  - Usability Rate (UR): Fraction of outputs with fewer than 3 “issues.”
  - Satisfaction Rate (SR): Fraction with zero issues.

### Quantitative Comparison

| Method        | PF (1–5) | SP (1–5) | DS (1–5) | Usability (%) |
|---------------|:--------:|:--------:|:--------:|:-------------:|
| DreamPoster   | 3.88     | 3.38     | 3.19     | 88.55         |
| GPT-4o        | 2.70     | 2.15     | 1.95     | 47.56         |
| SeedEdit 3.0  | 2.10     | 1.80     | 1.62     | 25.96         |

DreamPoster achieves the highest ratings in prompt compliance (PF), content (SP), and design (DS), and a usability rate of 88.6%. This is nearly double GPT-4o and triple SeedEdit 3.0’s usability rate. Qualitative analysis illustrates that DreamPoster preserves aspect-ratio flexibility and typographic hierarchy, whereas baselines distort content or lack layout sophistication.

## 5. Content Fidelity, Layout Flexibility, and Aesthetic Alignment

DreamPoster explicitly conditions on both user image and text, enabling faithful subject preservation and prompt following. Layout generation leverages enriched annotations (glyph and layout captions), supporting flexible aspect ratios and robust typographic hierarchy. The aesthetic alignment objective in Stage 3 directly optimizes for outputs that are aligned with human notions of design quality, as quantified via $s_{\rm aes}(\cdot)$. *This suggests* that the approach can generalize to a range of real-world poster design scenarios with minimal loss in content fidelity.

## 6. Applicability and Deployment

DreamPoster supports diverse forms of poster generation, including various compositional subtasks. The framework’s capacity for high-resolution, professional-quality outputs and high expert/user satisfaction underpins its deployment in production, notably in Bytedance’s Jimeng and other applications. The systematic annotation pipeline and progressive curriculum are directly transferable to related conditional image generation domains, particularly those demanding compositional control and aesthetic optimization [2507.04218].

Source: https://www.emergentmind.com/topics/dreamposter