Skip to content
Date and source

Dates follow the news timeline in Beijing time and include both endpoints.

Nov 16Mon
  1. Reddit StableDiffusion · 创作者与工作流分享16

    Minimax H3 Image Subject LoRA: Has anyone achieved good similarity?

    Training subject LoRA with Minimax H3 Image yields poor similarity, while the same dataset performs excellently on Qwen 2.1, Krea 2, and Moss. Increasing training epochs or steps showed no improvement; the Minimax H3 video model actually learns subjects better on static images. Motion learning is also difficult, and the poster is seeking advice on training settings for high-fidelity similarity.

  2. Reddit StableDiffusion · 创作者与工作流分享34

    How to get the highest quality video from Minimax H3 on a 4090? What about upscaling?

    Minimax H3 on a 4090 achieves best quality with settings of 1322x768, 24fps, 20 sampling steps, without turbo and without LoRA; the duration is usually about 4 seconds. For upscaling to 2k, you can use Regenerate 2k (Regen2k IC just released) or minimax_h3_latent_upscaler_3d_conv_v1_fp16. The author tested the API at 1080p and believes the quality improvement is not worth the price.

  3. WaveSpeedAI 官方博客69

    AI Video Character Consistency: A Single-Shot Workflow

    WaveSpeedAI's Dora shares an AI video character consistency shot workflow, using character reference kits, fixed identity details, and two-pass review to avoid character drift. Vidu 2.0 accepts 1-3 reference images, MiniMax H3 Highspeed accepts up to 9 reference images and 3 reference videos; when fixing drift, change only one input at a time.

    Why it matters: This workflow provides specific practices for character reference kits, fixed identity details, and two-pass review, and can be directly transferred to multi-shot AI video production.

  4. Reddit comfyui · 创作与工作流分享49

    Combining the LTX audio node with Minimax H3

    MiniMax H3 has poor lip-sync performance, but by using the LTX audio node to pack audio into the latent before generating video, you can get LTX-level lip-sync with H3 image quality. The workflow is simple: just add the LTX audio input node and merge the audio and visual latents. The author provides a workflow download and uses a VHS filter to process the video into an 80s VHS tape style.

  5. note · miyuu_2026 · 图像视频创作分享28

    11/16 [Wan2.2] Pink knit sweater girl street shooting episode

    The author used Krea2 to generate images, Wan2.2 14B to generate video, and Capcut for editing, producing an 18-second AI short drama of a girl in a pink knit sweater on a street shoot. 6 out of 8 shots were used. The production environment was Windows 11, RTX 5060 Ti 8GB, 32GB RAM, using ComfyUI workflow, resolution 1024x576px, 73 frames. The work consists of 3 seconds × 6 shots.

  6. Reddit StableDiffusion · 创作者与工作流分享56

    Minimax Image (Edit) Early Experience: Strong Concept-to-Photorealistic, Weak Multi-Reference Prompting

    Digital artist GrungeWerX tested Minimax Image (Edit) and found its concept-to-photorealistic capability outstanding, but multi-reference prompt following is weaker. He recommends starting at least at 16MP, with 30-50MP giving better results, and about 1 minute to generate 16MP on an RTX 3090 Ti.

  7. かみもと · 本地AI视频工作流72

    Hands-on test of the pDMD 2-step ComfyUI workflow for MiniMax-H3-Image: generation time on an RTX 4070 reduced by about 29%

    The pDMD 2-step LoRA for MiniMax-H3-Image, using ComfyUI 0.45.0 and an RTX 4070 12GB, reduced the average generation time for 9 image edits from 38.83 seconds to 27.62 seconds, a reduction of about 28.9%.

    Why it matters: The pDMD 2-step workflow reduced average generation time on an RTX 4070 from 38.83 seconds to 27.62 seconds. Composition and instruction following were largely maintained, but the lines became thicker.

  8. Reddit StableDiffusion · 创作者与工作流分享39

    OpenMontage Local Workflow Produced 2 Low-Cost Videos with Mediocre Results; Looking for Recommendations for More Realistic Image and Image-to-Video Models

    The author used OpenMontage to orchestrate a local workflow, combining gpt-image keyframes, Opus 5.5, and LTX-2.5 image-to-video generation to produce 2 low-cost videos. The two videos took about 6 hours including planning and rendering, but LTX-2.5 still had many artifacts, insufficient instruction following, and insufficient realism. The author asks for recommendations for image models that can compete with gpt-image and local image-to-video models that are better than LTX-2.5.

  9. Aitrepreneur · YouTube 创作视频59

    MiniMax H3 Image Local ComfyUI Workflow Tutorial: 4K Generation and Multi-Image Editing

    MiniMax H3 Image can run locally in ComfyUI, supporting native 4K image generation, editing existing images, and combining up to 16 reference images. After a week of hands-on testing, the author provides one-click installation options for local and RunPod setups, covering text-to-image generation, multi-reference editing, inpainting, outpainting, and a built-in prompt enhancer.

  10. Reddit comfyui · 创作与工作流分享12

    Why are most H3 LoRAs completely unusable?

    Most H3 LoRAs do not work properly: visual LoRAs show almost no difference, while action-related LoRAs fail completely. The poster says that the creative LoRAs they commonly use are mostly NSFW content and that they use pruned models and turbo LoRAs, suspecting that these two are the reason H3 LoRAs do not take effect.

  11. note · kongo_jun · 图像视频创作分享33

    Trying MiniMax-H3-Image-Restyle in ComfyUI

    Trying the MiniMax-H3-Image-Restyle style transfer LoRA in ComfyUI, with a generation time of 30 seconds and VRAM usage of 21GB. This LoRA needs to be used with MiniMax H3 Image's diffusion model, text encoder, VAE, and workflow files, and was tested on an RTX 4090 environment. The author believes that MiniMax H3 Image itself has strong image editing capabilities, and the LoRA can be used as a supplement when the effect is weak, but the actual results did not meet expectations.

Nov 15Sun
  1. Reddit StableDiffusion · 创作者与工作流分享71

    A single RTX 5090 running MiniMax H3 locally in ComfyUI completes a 2-minute 36-second sci-fi AI short drama.

    A creator in Budapest used a single RTX 5090 to run MiniMax H3 locally in ComfyUI, completing a 2-minute 36-second sci-fi AI short drama pilot episode. Each shot was first previewed at 672x384 with four steps for screening, then formally rendered at 1344x768 with twenty steps, downscaled to 1080p via 2x super-resolution, with no cloud usage and no per-use charges throughout.

    Why it matters: The author has published a two-stage output workflow for running MiniMax H3 locally on a single RTX 5090; the separation of preview and formal rendering can be migrated to other local video generation workflows.

  2. note · cocoredo_2509 · 图像视频创作分享47

    AI animation audio desync? Editing methods for natural integration of dialogue, sound effects, and BGM.

    For sound editing of AI-generated videos, first confirm the visual action, then adjust the synchronization of dialogue, sound effects, BGM, and subtitles. Check lip-sync with dialogue, sound effects with actions, ambient sound with scenes, and volume balance between BGM and dialogue, and pay attention to ambient sound transitions at scene changes. Finally, play back the entire video and confirm the exported file has no audio-visual desync.

  3. Reddit comfyui · 创作与工作流分享12

    Qwen image edit 2.1 malfunction

    When creating character sheets in ComfyUI, Qwen image edit 2.1 cannot maintain consistent character facial features; characteristics such as head length, head width, and freckles change. The user tried about 100 prompts, using the base extension and without enabling the prompt enhancer, but the problem remains unsolved. Their hardware is a 5070Ti 16GB plus another card (ComfyUI does not merge) and 64GB RAM.

  4. かみもと · 本地AI视频工作流66

    MiniMax H3 Image Hands-on Test: 10 Style Conversions Using Only Prompts

    The author tested MiniMax H3 Image, using an illustration as the source image, and completed 10 style conversions including watercolor, oil paint, pen, and woodblock print using only prompts; after enabling the prompt expansion built into the ComfyUI template, the style changes became more pronounced, but the average generation time increased from 60.50 seconds to 187.12 seconds.

    Why it matters: The author tested MiniMax H3 Image, completing 10 style conversions using only prompts, and compared the effects and time costs of the prompt expansion toggle, providing a reference for local image stylization.

  5. SuSu_酥酥73

    MiniMax H3 Image quantized version reduces VRAM from 68GB to 6.4GB; 8GB graphics cards can run 4K image generation.

    The open-source community has quantized MiniMax H3 Image to Q2_K, reducing VRAM requirements from 68GB to 6.4GB, allowing 8GB graphics cards to run 4K image generation locally. The quantized version natively outputs 4096×4096 and 21:9 widescreen, official distillation produces images in 6 steps, integrates generation and editing, and supports up to 16 reference images.

    Why it matters: The quantized version reduces VRAM requirements from 68GB to 6.4GB, allowing 8GB graphics card users to run 4K image generation locally, which can be used to evaluate whether to replace existing image generation workflows.

  6. Reddit StableDiffusion · 创作者与工作流分享53

    H3 images rough? Three settings to fix it.

    The author shares three settings to fix H3 image roughness: increase size from default 4MP to 16MP (3552x4736), switch sampler to SamplerLCM with noise 0.9, and write a long prompt describing camera, lighting, and skin texture. The workflow file is in the first comment.

  7. Monster社区 · 中文ComfyUI实测与工作流45

    How to choose between Claude and Seedance: A comparison method using five video tasks

    T8star-Aix compared two approaches—Claude's code-rendered animation and Seedance's direct video generation—using five video tasks. The tasks covered explaining the principle of heart pumping, a cat knocking over a water glass, verifying numbers in a bar chart, an elderly person's emotions at a station, and a local modification test. When clear explanations, precise text/data, or repeated local adjustments are needed, prioritize evaluating code-rendered animation; when realistic materials, complex lighting and shadows, or expressive human faces are needed, prioritize evaluating generated video. For retesting, record your own actual time and cost, and do not copy the demonstration numbers as performance promises.

  8. Monster社区 · 中文ComfyUI实测与工作流54

    MiniMax H3 and MOSS Image Generation: Model Comparison and Accelerated LoRA Trade-offs

    Author T8star-Aix compares MiniMax H3 and MOSS image generation models, showcasing ComfyUI workflows and accelerated LoRA trade-offs. The article provides a checking sequence: establish a baseline with the same prompt, distinguish style from structure, then confirm whether the acceleration gains are worth the detail loss. It also notes that 6 steps, 15 steps, and 8 steps cannot be blindly copied without considering the model.

  9. note · akb428 · 图像视频创作分享29

    Create beautiful girl character parties with Anima and Pixubus 4: usable for novels, event CG, and animation

    Pixubus 4 and Anima can one-click generate a party of 1-5 beautiful girl characters, supports Japanese prompts, and no English tags are required. The party can be imported into Inkubus v4.5 for novel writing and output as EPUB, or via Vidubus 1.5 generate a 15-second animation with sound. Anima was developed by CircleStone Labs and Comfy Org, trained on millions of anime images and approximately 800,000 art images.

Nov 14Sat
  1. Reddit comfyui · 创作与工作流分享20

    MOSS Image excels in semi-realistic anime style.

    Reddit user Phenilethylamine showcased semi-realistic anime works generated using the MOSS-Image-SFT_clear model with LoRA. The workflow uses layer skip 22, interval 0-0.3, paired with two LoRAs: Mixdark Style and Realistic Background Enchanted, and upscaled with SEEDVR2.

  2. Reddit StableDiffusion · 创作者与工作流分享60

    Minimax-H3-Image can use a marked map as a reference when generating images.

    Minimax-H3-Image can use a marked map as a reference image when generating images, to control camera position and character placement. The author used a top-down layout diagram to mark camera arrows and character positions, and combined multiple reference images to generate cinematic stills; generation on a 5090M (24GB) took about 300 seconds, and the time increases with the number of reference images.

  3. Monster社区 · 中文ComfyUI实测与工作流39

    Seedream 5.5: ComfyUI Image Generation, Reference Editing, and Layering Tutorial

    Seedream 5.5 can follow three routes in ComfyUI: text-to-image generation, reference-image editing, and poster layering. The demonstration outputs 720×1280 PNG and emphasizes first distinguishing t2i and i2i tasks. Reference editing requires checking the face, outline, and pose, while poster layering requires checking text, edges, and inpainted areas.

  4. Reddit comfyui · 创作与工作流分享67

    Making a 10-minute dialogue-free animated film in ComfyUI using MiniMax H3, LTX-2.5, and Wan 2.2; the author shares problems encountered and solutions.

    The author used MiniMax H3, LTX-2.5, and Wan 2.2 in ComfyUI to complete a 9:51 dialogue-free animated short film 'DAY 731', with most shots generated by H3.

    Why it matters: The author publicly shared parameters and pitfalls-to-avoid experience for making a long-form film on a 3060 12GB, including specific issues such as resolution, sampler, and VAE redraw.

  5. Luma News44

    Recommended Native Audio AI Video Generators: Luma and Google Veo

    Native audio AI video generators can synchronously output dialogue, sound effects, and ambient sound in a single generation; Luma and Google Veo are representative platforms. Luma's Ray 3.2 supports multiple keyframes, HDR, EXR export, and Motion Transfer, and uses Uni-1 to maintain project context. Google Veo is positioned as an enterprise solution with broad platform access capabilities.

  6. Synthesia · 官方博客22

    Synthesia: How to Speed Up eLearning Content Production (While Maintaining Quality)

    Synthesia recommends using a unified intake form, prioritizing requests by scores for business impact, evidence, workload, and urgency, and using AI drafting to speed up eLearning content production while maintaining quality. Its 2027 AI in L&D report shows that 75% of learning professionals believe AI leads to faster content creation, but only 11% mention improved job performance and 35% continuously measure it.

  7. Reddit StableDiffusion · 创作者与工作流分享39

    MiniMax H3 Image's editing at 60MP is astonishing: from concept to realism

    MiniMax H3 Image edits concept art into photorealistic photos at 60MP resolution, with astonishing detail. The test prompt required preserving subject similarity, pose, and background, while improving skin, hair, fabric texture, background details, and lighting. Generating a 60MP image on a 4090 takes about 2 minutes.

  8. Reddit comfyui · 创作与工作流分享61

    Hands-on test and sharing of sync.so's ComfyUI lip-sync workflow

    The author tested sync.so's ComfyUI lip-sync workflow and shared installation steps, model prices, and duration limits. Two usage methods: built-in nodes and custom nodes. lipsync-2 is about $0.05/second, free quota is 3 times per month for 20 seconds. In the test, only the speaking actor's mouth was reanimated; the rest of the footage and the reference dialogue were unaffected.

  9. Reddit StableDiffusion · 创作者与工作流分享18

    Has anyone tried building a one-click AI short film with ComfyUI?

    A Reddit user attempted to build a one-click AI short film workflow with ComfyUI, using Qwen 2.1 for image generation and face swapping, MiniMax H3 for video generation, and LTX 2.5 for video upscaling. The workflow runs normally, but the output is still 'AI SLOP'. The author believes the problem lies in the prompts, and used Claude AI to create nodes that automate the three workflows into one click.

  10. Reddit StableDiffusion · 创作者与工作流分享38

    Experience running H3 Image on an RTX 2060 Super

    On an RTX 2060 Super, H3 Image can generate 16mp images, taking about 2.5 minutes without reference images. When using 3 reference images, each 16mp image takes about 10 minutes; larger reference images provide more detail but take longer. This model runs smoothly on the RTX 2060 Super, whereas other models run out of memory or take too long at resolutions above 2mp or 4mp.

Nov 13Fri
  1. Reddit StableDiffusion · 创作者与工作流分享62

    RTX 5070 Ti ComfyUI benchmark: overclocking reduces generation time by about 10%, tuned parameters cut H3 video time by more than half

    The author tested the effects of GPU overclocking, ComfyUI startup parameters, and connecting the display to the iGPU on generation speed on an RTX 5070 Ti. Overclocking consistently reduced generation time by about 9-12%, and tuned parameters cut H3 video time by more than half.

  2. Higgsfield 官方博客67

    How to Create AI Videos Directly in ChatGPT (Complete Guide)

    Creating AI videos in ChatGPT requires connecting video apps through the plugin directory. This guide covers the categories and billing rules of 17 plugins, including Higgsfield, Runway, HeyGen, and invideo.

    Why it matters: This guide compares the billing rules and capability differences of 17 video apps in the ChatGPT plugin directory, and provides reusable prompt templates to help creators choose tools and control costs.

  3. Reddit StableDiffusion · 创作者与工作流分享15

    How to maintain facial consistency in MiniMax H3 video during dynamic camera motion?

    When users generate dynamic camera motion videos with MiniMax H3 video, the face and eyes exhibit slight flickering/jitter. The configuration is a single-chain exact latent continuation in ComfyUI at 1344×768 @ 24 fps, 192 frames (8 seconds), using prompt + start_image + 2 reference images (Qwen3-VL encoder), sampler res_multistep 32 steps, cfg 1, fixed seed, but the jitter still persists.

  4. Leeveo AI Tutorials · 创作教程与项目分享39

    Stop randomly generating AI images; use GPT to build a complete cinematic world.

    GPT image generation can be used to build a cinematic environment reference sheet, rather than generating random images one by one. The reference sheet covers camera positions, lighting, key locations, visual assets, and style directions, and can support image-to-video workflows such as Kling and Veo. The same environment can be reused across multiple projects, such as short films, game concepts, and animation series.