1 00:00:03,030 --> 00:00:05,400 Evan: Welcome to Daily Paper Cast. 2 00:00:05,448 --> 00:00:12,648 Ashley: Today's paper is from the Hugging Face daily paper list of October 20, 2026, with 30 upvotes. 3 00:00:12,696 --> 00:00:20,416 Evan: We're discussing 'MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion.' 4 00:00:20,472 --> 00:00:24,172 Ashley: The first two authors are Houyuan Chen and Fanzhou Wang. 5 00:00:24,172 --> 00:00:28,312 The corresponding author is Hao Zhao, from SenseTime Research. 6 00:00:28,408 --> 00:00:31,828 Evan: Alright, let's dive right into the Introduction. 7 00:00:31,828 --> 00:00:34,308 Can you kick us off with the background, Ashley? 8 00:00:34,368 --> 00:00:40,828 Ashley: Multi-turn image editing is all about progressively modifying an image through a sequence of user instructions. 9 00:00:40,828 --> 00:00:50,808 Unlike single-turn editing, where changes are made in one shot, multi-turn editing allows users to refine visual content incrementally as their creative intentions evolve. 10 00:00:50,856 --> 00:00:56,276 Evan: That makes sense, especially for more complex edits where precision is key. 11 00:00:56,276 --> 00:00:59,696 But what challenges does this process face? 12 00:00:59,760 --> 00:01:01,020 Ashley: Exactly, Evan. 13 00:01:01,020 --> 00:01:12,190 The main challenges are twofold: first, the stability of visual quality, which often deteriorates with each successive edit, leading to accumulating artifacts and loss of fine-grained details. 14 00:01:12,190 --> 00:01:19,770 Second, maintaining image content consistency, where unintended changes can occur in preserved regions or with prior edits. 15 00:01:19,770 --> 00:01:28,280 These problems are largely due to exposure bias, which is the mismatch between training conditions and the actual settings encountered during inference. 16 00:01:28,344 --> 00:01:29,084 Evan: Got it. 17 00:01:29,084 --> 00:01:32,744 So, what's the proposed solution in this paper? 18 00:01:32,808 --> 00:01:38,028 Ashley: To tackle these issues, the authors propose a training paradigm called MT Forcing. 19 00:01:38,028 --> 00:01:46,308 The core idea is to leverage Single-Turn Distribution Matching Distillation, or Single-Turn DMD, and Teacher-Forcing Guidance. 20 00:01:46,308 --> 00:01:54,478 Single-Turn DMD supervises student self-rollouts with a teacher model that applies cumulative edits to the source image in a single turn. 21 00:01:54,478 --> 00:01:58,968 This helps maintain visual quality without recursive error accumulation. 22 00:01:59,016 --> 00:02:00,486 Evan: Interesting. 23 00:02:00,486 --> 00:02:04,976 How does this Single-Turn DMD actually work in practice? 24 00:02:05,070 --> 00:02:06,130 Ashley: Good question. 25 00:02:06,130 --> 00:02:14,260 Essentially, it reorganizes all instructions up to the current turn into a bundled prompt that expresses them as a single edit to the source image. 26 00:02:14,260 --> 00:02:20,460 The teacher executes this bundled edit in one go, providing a high-quality reference for the student model. 27 00:02:20,460 --> 00:02:30,160 The student, on the other hand, performs autoregressive self-rollouts during training, meaning it generates each edit based on its previous imperfect outputs. 28 00:02:30,160 --> 00:02:38,320 A loss function then aligns the student's output distribution with the teacher's single-turn output distribution for the same cumulative intent. 29 00:02:38,376 --> 00:02:47,976 Evan: But this doesn't entirely solve the issue of cross-turn consistency, does it? 30 00:02:48,024 --> 00:02:49,454 Ashley: You're spot on, Evan. 31 00:02:49,454 --> 00:02:52,064 That's where Teacher-Forcing Guidance comes in. 32 00:02:52,064 --> 00:02:56,564 This method supervises history-dependent transitions on reference trajectories. 33 00:02:56,564 --> 00:03:10,844 It's applied in two stages: first, as a warmup to teach the model to produce consistent edits based on reference histories, and second, during the Single-Turn DMD training as a regularizer to preserve cross-turn content consistency. 34 00:03:10,956 --> 00:03:22,616 Evan: So, basically, MT Forcing combines these two strategies to ensure stable visual quality and consistent content across multiple editing turns. 35 00:03:22,680 --> 00:03:23,620 Ashley: Exactly. 36 00:03:23,620 --> 00:03:32,170 The complementary nature of Single-Turn DMD and Teacher-Forcing Guidance aims to address the inherent issues of multi-turn image editing. 37 00:03:32,170 --> 00:03:43,820 The authors conducted extensive quantitative and qualitative experiments to demonstrate that their approach significantly improves the multi-turn editing capabilities of unified multimodal models. 38 00:03:43,872 --> 00:03:46,372 Evan: Well, that's the Introduction section. 39 00:03:46,372 --> 00:03:50,492 Stay tuned as we dive deeper into the methods and experiments next. 40 00:03:51,806 --> 00:03:56,746 Evan: Alright, let's delve into the heart of the paper — the Method section. 41 00:03:56,746 --> 00:04:02,826 Ashley, can you start us off with how the authors define the multi-turn editing problem? 42 00:04:02,882 --> 00:04:03,932 Ashley: Certainly. 43 00:04:03,932 --> 00:04:10,102 The authors frame a multi-turn editing session as consisting of a source image and a series of instructions. 44 00:04:10,102 --> 00:04:20,352 For each turn, the model needs to generate an edited image based on the current instruction and the accumulated history of previous instructions and their corresponding outputs. 45 00:04:20,352 --> 00:04:28,222 This is expressed mathematically in the paper, where the model generates each output image by conditioning on this cumulative context. 46 00:04:28,334 --> 00:04:36,534 Evan: How do they align the training and inference processes to handle this cumulative context? 47 00:04:36,578 --> 00:04:38,618 Ashley: That alignment is crucial. 48 00:04:38,618 --> 00:04:46,418 During training, the model often conditions on clean reference histories, while during inference, it relies on its own imperfect outputs. 49 00:04:46,418 --> 00:04:49,268 This mismatch is known as exposure bias. 50 00:04:49,268 --> 00:04:56,738 To bridge this gap, the authors propose Single-Turn Distribution Matching Distillation, or Single-Turn DMD. 51 00:04:56,906 --> 00:05:03,536 Evan: Tell us more about this Single-Turn DMD approach. 52 00:05:03,536 --> 00:05:05,146 How does it work? 53 00:05:05,210 --> 00:05:12,810 Ashley: Single-Turn DMD works by reorganizing all editing instructions up to the current turn into what's known as a bundled prompt. 54 00:05:12,810 --> 00:05:18,140 This bundled prompt is essentially a single, comprehensive edit applied to the source image. 55 00:05:18,140 --> 00:05:27,110 The teacher model executes this bundled edit in one go, providing high-quality reference outputs without the risk of error accumulation across turns. 56 00:05:27,230 --> 00:05:34,230 Evan: So, this bundled prompt represents the cumulative editing intent as a single action, correct? 57 00:05:34,274 --> 00:05:35,304 Ashley: Exactly. 58 00:05:35,304 --> 00:05:49,544 The authors illustrate this with examples, like combining 'add a vase' with 'make it blue' into a single prompt: 'add a blue vase.' The teacher model, conditioned on this prompt, provides a clear reference by generating the image in one turn. 59 00:05:49,544 --> 00:05:58,114 The student model, meanwhile, performs autoregressive self-rollouts, generating an edit based on its previous outputs, which might not be perfect. 60 00:05:58,248 --> 00:06:02,378 Evan: And how does the student model learn from these rollouts? 61 00:06:02,426 --> 00:06:03,696 Ashley: Great question. 62 00:06:03,696 --> 00:06:11,916 The student’s performance is evaluated using a loss function that aligns its output distribution with the teacher's single-turn output distribution. 63 00:06:11,916 --> 00:06:22,606 They call this the Distribution Matching Distillation loss, or DMD loss, which trains the student to maintain visual quality similar to the reference provided by the teacher model. 64 00:06:22,718 --> 00:06:29,398 Evan: But even with this mechanism, maintaining consistency across multiple turns is still a challenge, right? 65 00:06:29,450 --> 00:06:30,260 Ashley: Correct. 66 00:06:30,260 --> 00:06:32,860 That's where the Teacher-Forcing Guidance comes in. 67 00:06:32,860 --> 00:06:41,740 To boost cross-turn consistency, this method supervises the history-dependent transitions by training the student on consistent reference trajectories. 68 00:06:41,740 --> 00:06:45,590 This is done in two steps: warmup and regularization. 69 00:06:45,650 --> 00:06:49,330 Evan: Can you break down these two stages for us? 70 00:06:49,394 --> 00:06:50,124 Ashley: Of course. 71 00:06:50,124 --> 00:06:55,234 In the warmup stage, the pretrained model is fine-tuned using reference editing trajectories. 72 00:06:55,234 --> 00:07:00,684 This helps the student model learn to produce consistent edits based on coherent reference histories. 73 00:07:00,684 --> 00:07:15,154 During the actual Single-Turn DMD training, Teacher-Forcing Guidance is applied as a regularizer with a lower weight to ensure that history-dependent behaviors are preserved, which aids in maintaining content consistency across multiple editing turns. 74 00:07:15,218 --> 00:07:30,318 Evan: So, to summarize, the warmup stage educates the model on producing consistent edits, and during the main training, this guidance acts like a constraint to maintain that learned consistency. 75 00:07:30,362 --> 00:07:31,412 Ashley: Precisely. 76 00:07:31,412 --> 00:07:45,082 By combining these two strategies, Single-Turn DMD for quality and Teacher-Forcing Guidance for consistency, the authors claim that MT Forcing significantly enhances the stability and accuracy of multi-turn image editing. 77 00:07:45,196 --> 00:07:47,176 Evan: It sounds comprehensive. 78 00:07:47,176 --> 00:07:52,486 But how do they practically apply this MT Forcing approach in their experiments? 79 00:07:52,538 --> 00:08:02,948 Ashley: In their experiments, they use two unified multimodal models, SenseNova-U1 and SenseNova-U1.5, to validate the effectiveness of MT Forcing. 80 00:08:02,948 --> 00:08:07,038 They conduct extensive quantitative and qualitative evaluations. 81 00:08:07,142 --> 00:08:11,822 Evan: And what kind of metrics do they use to measure performance? 82 00:08:11,882 --> 00:08:20,642 Ashley: They utilize metrics like Instruction Following (IF), Content Consistency (CC), and Drift Robustness Score (DRS). 83 00:08:20,642 --> 00:08:33,102 IF measures how well the model follows the current instruction, CC evaluates how well the content is preserved across multiple turns, and DRS assesses the model's robustness against visual quality degradation. 84 00:08:33,146 --> 00:08:34,826 Evan: Interesting. 85 00:08:34,826 --> 00:08:38,946 And did they introduce any new metrics? 86 00:08:39,002 --> 00:08:46,422 Ashley: Yes, they introduced a new metric called Pixel, which measures the geometric alignment between successive editing outputs. 87 00:08:46,422 --> 00:08:55,422 Using this metric, they quantify the mean matched-point displacement in pixels between consecutive images to evaluate geometric consistency. 88 00:08:55,466 --> 00:09:07,286 Evan: So, with these metrics, how did MT Forcing perform compared to the baseline models? 89 00:09:07,346 --> 00:09:13,566 Ashley: The results demonstrated that MT Forcing outperformed the baseline models in all key areas. 90 00:09:13,566 --> 00:09:21,926 It showed significant improvements in instruction following, content consistency, visual quality, and geometric alignment. 91 00:09:22,046 --> 00:09:25,876 Evan: Alright, that covers the Method section. 92 00:09:25,876 --> 00:09:31,846 Stick around as we move on to the Experiments and Results next. 93 00:09:33,129 --> 00:09:37,299 Evan: Now let's dive into the Experiments and Results section. 94 00:09:37,299 --> 00:09:42,679 Ashley, how did the authors set up their experiments to evaluate MT Forcing? 95 00:09:42,723 --> 00:09:51,643 Ashley: The authors evaluated MT Forcing on two unified multimodal models, SenseNova-U1 and SenseNova-U1.5. 96 00:09:51,643 --> 00:09:57,883 They conducted extensive experiments using these models to benchmark performance across various metrics. 97 00:09:57,939 --> 00:10:01,819 Evan: What metrics did they use to measure the performance? 98 00:10:01,875 --> 00:10:11,735 Ashley: They used several key metrics: Instruction Following, or IF; Content Consistency, or CC; and Drift Robustness Score, or DRS. 99 00:10:11,735 --> 00:10:14,915 IF measures how well the current instruction is followed. 100 00:10:14,915 --> 00:10:25,455 CC assesses the preservation of cumulative content across all instructions, and DRS evaluates how well the model maintains visual quality over multiple turns. 101 00:10:25,515 --> 00:10:30,315 Evan: Were there any new metrics introduced in their evaluation? 102 00:10:30,363 --> 00:10:37,413 Ashley: Yes, they introduced a metric called Pixel, which quantifies the geometric alignment between successive images. 103 00:10:37,413 --> 00:10:44,403 It measures the mean matched-point displacement in pixels to assess how well geometric consistency is preserved. 104 00:10:44,511 --> 00:10:45,791 Evan: Interesting. 105 00:10:45,791 --> 00:10:51,051 How did MT Forcing fare with these metrics compared to the baseline models? 106 00:10:51,099 --> 00:10:52,849 Ashley: The results were impressive. 107 00:10:52,849 --> 00:10:57,129 MT Forcing outperformed the baseline models across all key metrics. 108 00:10:57,129 --> 00:11:04,559 For example, their experiments showed significant gains in instruction following, content consistency, and visual quality. 109 00:11:04,681 --> 00:11:07,491 Evan: Can you give us a breakdown of these improvements? 110 00:11:07,569 --> 00:11:08,549 Ashley: Definitely. 111 00:11:08,549 --> 00:11:10,669 Let's look at Table 1 from the paper. 112 00:11:10,669 --> 00:11:22,839 For the SenseNova-U1 model, from turn 4 onwards, MT Forcing improved Instruction Following by about 0.83 points at turn 2 and by as much as 7.45 points by turn 8. 113 00:11:22,839 --> 00:11:29,349 Content Consistency also saw significant enhancements, with an increase of up to 6.82 points. 114 00:11:29,349 --> 00:11:36,539 Similarly, Drift Robustness Score increased consistently from 4.07 to 4.22 at turn 8. 115 00:11:36,539 --> 00:11:44,699 Finally, Pixel measures showed a dramatic reduction in geometric misalignment, bringing it down by approximately 88.4%. 116 00:11:44,763 --> 00:11:47,033 Evan: That's quite an improvement! 117 00:11:47,033 --> 00:11:49,563 How about the qualitative results? 118 00:11:49,611 --> 00:11:53,961 Ashley: Qualitative comparisons further highlight the effectiveness of MT Forcing. 119 00:11:53,961 --> 00:12:01,741 Figures in the paper show that the MT Forcing models preserved textures and maintained a coherent scene structure better than the baseline models. 120 00:12:01,741 --> 00:12:16,111 For instance, in one set of images, MT Forcing maintained the integrity of details like the red car body and roof box consistently across multiple editing turns, whereas the baseline models exhibited oversaturation and harsh textures. 121 00:12:16,155 --> 00:12:24,315 Evan: So, visually, the MT Forcing approach produced cleaner and more consistent results across multiple edits. 122 00:12:24,363 --> 00:12:25,633 Ashley: Exactly, Evan. 123 00:12:25,633 --> 00:12:34,543 Additionally, they conducted ablation studies to quantify the contributions of Single-Turn DMD and Teacher-Forcing Guidance separately and combined. 124 00:12:34,543 --> 00:12:40,423 The results illustrated that both components independently improved performance compared to the baseline. 125 00:12:40,423 --> 00:12:46,003 However, the best results were achieved when both components were combined in MT Forcing. 126 00:12:46,059 --> 00:12:50,739 Evan: Can you give us more specifics on the ablation studies? 127 00:12:50,787 --> 00:12:51,547 Ashley: Sure. 128 00:12:51,547 --> 00:12:56,717 Table 2 in the paper shows eight-turn mean scores across metrics for various combinations. 129 00:12:56,717 --> 00:13:05,787 For example, using Single-Turn DMD alone improved the Drift Robustness Score to 4.12 from a baseline of 3.83. 130 00:13:05,787 --> 00:13:13,037 Adding Teacher-Forcing Guidance alone improved Instruction Following to 9.10 from the baseline 8.47. 131 00:13:13,037 --> 00:13:26,727 Combining both methods yielded the highest improvements: an IF score of 9.66, CC score of 8.40, DRS score of 4.30, and Pixel score reduced to 0.74. 132 00:13:26,837 --> 00:13:34,447 Evan: It's clear that both components of MT Forcing are essential for achieving these superior results. 133 00:13:34,447 --> 00:13:39,767 Were there any experiments related to the regularization weight in the Teacher-Forcing Guidance? 134 00:13:39,819 --> 00:13:44,629 Ashley: Yes, the authors explored the impact of different regularization weights. 135 00:13:44,629 --> 00:13:55,329 They tested weights of 1, 0.5, 0.25, and 0.1 and found that higher weights led to visible blurring in the output images. 136 00:13:55,329 --> 00:14:07,699 Reducing the weight to 0.25 or 0.1 restored finer details and sharper boundaries, indicating a favorable balance between content consistency and visual clarity. 137 00:14:07,755 --> 00:14:09,705 Evan: That's quite thorough. 138 00:14:09,705 --> 00:14:19,685 So to wrap it up, the extensive experiments demonstrated not only the quantitative but also qualitative improvements brought by MT Forcing. 139 00:14:19,685 --> 00:14:24,155 And that concludes our talk on the Experiments and Results section of the paper. 140 00:14:24,219 --> 00:14:27,619 Ashley: Stay tuned as we move on to the Related Work next. 141 00:14:28,877 --> 00:14:32,667 Evan: Now it's time to delve into the Related Work section. 142 00:14:32,667 --> 00:14:38,717 Ashley, how do previous endeavors compare to the approach laid out in this paper? 143 00:14:38,765 --> 00:14:49,085 Ashley: The paper provides a comprehensive overview of previous work in three main areas: multi-turn image editing, diffusion distillation, and autoregressive diffusion. 144 00:14:49,085 --> 00:14:51,665 Let's start with multi-turn image editing. 145 00:14:51,825 --> 00:14:57,425 Evan: How has multi-turn image editing evolved over time? 146 00:14:57,485 --> 00:15:06,075 Ashley: Well, traditional instruction-based image editing approaches, like those mentioned in the paper, typically apply single edits in isolation. 147 00:15:06,075 --> 00:15:12,925 In multi-turn settings, users give multiple instructions in sequence, allowing them to refine the image progressively. 148 00:15:12,989 --> 00:15:19,009 Evan: So, how have recent approaches tried to address the issues in multi-turn editing? 149 00:15:19,061 --> 00:15:29,071 Ashley: Recent methods extend the capabilities of unified multimodal models by incorporating prior instructions and generated images into a shared context for successive edits. 150 00:15:29,071 --> 00:15:41,461 For example, VINCIE uses in-context editing from video-derived sequences, and AnchorEdit introduces causal self-rollout training and memory mechanisms to preserve identity and context across turns. 151 00:15:41,525 --> 00:15:42,715 Evan: Interesting. 152 00:15:42,715 --> 00:15:45,225 What are some other techniques mentioned? 153 00:15:45,269 --> 00:15:51,749 Ashley: Techniques like FreqEdit focus on preserving high-frequency features to reduce degradation over turns. 154 00:15:51,749 --> 00:15:56,689 MT-EditFlow uses reinforcement learning to optimize editing trajectories. 155 00:15:56,741 --> 00:16:02,621 Evan: It seems like there's a lot of innovation in dealing with the complexities of multi-turn edits. 156 00:16:02,621 --> 00:16:03,801 What's next? 157 00:16:03,845 --> 00:16:06,825 Ashley: The next area is diffusion distillation. 158 00:16:06,825 --> 00:16:14,405 This is all about reducing the computation required for iterative diffusion sampling while maintaining high generation quality. 159 00:16:14,453 --> 00:16:19,973 Evan: How does diffusion distillation contribute to this field? 160 00:16:20,051 --> 00:16:28,021 Ashley: Diffusion distillation methods, like the teacher-trajectory regression, learn sampling dynamics to reduce noise while maintaining clarity. 161 00:16:28,021 --> 00:16:34,321 Consistency-based models ensure that predictions along a diffusion trajectory converge to a common end-point. 162 00:16:34,373 --> 00:16:40,793 Evan: And this consistency is crucial for maintaining high visual quality across multiple turns, right? 163 00:16:40,853 --> 00:16:41,773 Ashley: Exactly. 164 00:16:41,773 --> 00:16:50,073 Methods such as consistency-based interval mappings and transitive supervision have been explored to enhance stability and scalability. 165 00:16:50,073 --> 00:16:56,813 These techniques help ensure the generated output remains consistent and high-quality over multiple iterations. 166 00:16:56,861 --> 00:17:04,341 Evan: Could you mention some specific works in diffusion distillation? 167 00:17:04,427 --> 00:17:05,277 Ashley: Of course. 168 00:17:05,277 --> 00:17:16,197 Notable works include InstaFlow, which simplifies diffusion steps to a single consistent path, and rCM, which combines distillation with score regularization for improved results. 169 00:17:16,197 --> 00:17:24,597 Other studies, like SwiftBrush and Phased DMD, have also shown progress in improving the efficiency and quality of diffusion-based models. 170 00:17:24,653 --> 00:17:28,003 Evan: Great, that brings a lot of context. 171 00:17:28,003 --> 00:17:30,633 What about autoregressive diffusion? 172 00:17:30,717 --> 00:17:39,877 Ashley: Autoregressive diffusion is pivotal because it generates content in a sequential, causal manner, which is essential for tasks like multi-turn image editing. 173 00:17:39,877 --> 00:17:50,517 Unified multimodal models, or UMMs, extend sequence modeling to visual tasks by combining autoregressive context modeling with diffusion-based synthesis. 174 00:17:50,573 --> 00:17:55,793 Evan: So, how do these models integrate with the techniques we've just discussed? 175 00:17:55,853 --> 00:18:01,273 Ashley: These models build on the principles of bidirectional and few-step autoregressive generation. 176 00:18:01,273 --> 00:18:13,173 For instance, methods like Self Forcing and Diffusion Forcing train on self-generated rollouts to minimize discrepancies between training and inference, aligning well with MT Forcing's approach. 177 00:18:13,229 --> 00:18:18,809 Evan: But there have to be some limitations to these approaches, right? 178 00:18:18,869 --> 00:18:27,319 Ashley: One challenge is transferring these autoregressive methods directly to multi-turn image editing without introducing errors over long horizons. 179 00:18:27,319 --> 00:18:39,689 Streaming methods and long-range context modeling, like in CausVid, and error control techniques, help, but they don't fully resolve the issues of consistency and stability over many editing turns. 180 00:18:39,749 --> 00:18:49,109 Evan: So the paper builds on these methods, addressing their limitations with MT Forcing's novel approach to training and consistency? 181 00:18:49,157 --> 00:18:50,307 Ashley: Exactly. 182 00:18:50,307 --> 00:19:03,137 MT Forcing leverages these prior works by integrating Single-Turn DMD and Teacher-Forcing Guidance, tackling stability and consistency issues more effectively across various turns and instructions. 183 00:19:03,267 --> 00:19:06,297 Evan: And that wraps up the Related Work section. 184 00:19:06,297 --> 00:19:10,837 Stay with us as we move forward to the Conclusion next. 185 00:19:12,102 --> 00:19:17,242 Evan: Now let's summarize the key contributions and takeaways from this paper. 186 00:19:17,242 --> 00:19:20,102 Ashley, what stands out the most to you? 187 00:19:20,166 --> 00:19:28,186 Ashley: The paper introduces MT Forcing, a novel approach designed to improve the stability and consistency of multi-turn image editing. 188 00:19:28,186 --> 00:19:36,666 The core innovations include Single-Turn Distribution Matching Distillation, or Single-Turn DMD, and Teacher-Forcing Guidance. 189 00:19:36,726 --> 00:19:41,266 Evan: Can you briefly recap how these methods work together? 190 00:19:41,340 --> 00:19:42,330 Ashley: Of course. 191 00:19:42,330 --> 00:19:48,900 Single-Turn DMD works by reorganizing all editing instructions up to the current turn into a bundled prompt. 192 00:19:48,900 --> 00:19:55,230 This prompt represents the cumulative editing intent as a single action, applied by a teacher model in one go. 193 00:19:55,230 --> 00:19:58,930 This helps maintain visual quality without error accumulation. 194 00:19:58,930 --> 00:20:11,430 Meanwhile, Teacher-Forcing Guidance supervises history-dependent transitions, first through a warmup to teach the model consistent editing, and then during training as a regularizer to preserve cross-turn content consistency. 195 00:20:11,478 --> 00:20:21,138 Evan: So, by combining these methods, MT Forcing addresses the key issues of stability and content consistency in multi-turn image editing. 196 00:20:21,198 --> 00:20:22,158 Ashley: Exactly. 197 00:20:22,158 --> 00:20:35,118 The extensive quantitative and qualitative experiments demonstrate significant improvements across multiple metrics, including instruction following, content consistency, drift robustness, and geometric alignment. 198 00:20:35,166 --> 00:20:48,246 Evan: That wraps up our discussion on 'MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion.' It's clear this paper offers substantial advancements in the field. 199 00:20:48,294 --> 00:20:51,084 Ashley: Thank you for tuning in to Daily Paper Cast. 200 00:20:51,084 --> 00:20:54,714 We hope you found today's episode as insightful as we did. 201 00:20:54,774 --> 00:21:02,774 Evan: Join us next time as we explore more cutting-edge research from the world of AI, NLP, computer vision, and beyond. 202 00:21:02,774 --> 00:21:05,294 Don't forget to subscribe to stay updated. 203 00:21:05,294 --> 00:21:08,314 Until then, keep learning and innovating!