Flux 2 AI Model – Better Than Qwen Edit?
Takeaways
- 🚀Flux 2 has officially released, bringing major excitement to the creator. This powerful tool, Flux 2 AI, opens up new possibilities for creative workflows.
- 🖼️ The model specializes in hyper-realism and offers strong image-editing capabilities comparable to Qwen Edit.
- 🔧 It supports reference photos and image editing with noticeably improved performance.
- ⏱️ The creator has only tested it for about five minutes but is already impressed with the results.
- 💾 An FP8 example workflow is available for free on the creator’s Patreon for users with lower VRAM needs.
- 💻 Despite FP8, the model still requires significant hardware—ideally 24GB+ VRAM; the creator runs it on a 5090 with 32GB.
- ⚠️ The workflow caused some crashes and high RAM usage, so optimization and feedback are still needed.
- 🐢 Flux 2 runs about three times slower per iteration than Flux Korea (around 6.5 seconds per iteration).
- 📸 Early test generations show promising quality even with simple prompts and no LoRAs.
- 🛠️ The creator plans to train LoRAs for Flux 2 and update their LoRA training tool within 1–2 weeks.
- 🤝 A free community Discord is launching soon for users to chat, learn, and share feedback.
Q & A
What is the main topic of the video?
-The video discusses the release of the Flux 2 AI model and compares its capabilities to Qwen Edit.
Why is the creator excited about Flux 2?
-The creator is excited because Flux 2 focuses on hyper-realism and offers improved image editing and reference-based generation.
How does Flux 2 compare to Qwen Edit?
-Flux 2 is positioned as a competitor to Qwen Edit, offering similar image editing features but with noticeable improvements.
How long has the creator been testing Flux 2?
-The creator has only been testing Flux 2 for about five minutes before recording the video.
What hardware does the creator use to run Flux 2?
-The creator is using an NVIDIA 5090 GPU with 32GB of VRAM.
Is Flux 2 resource-intensive?
-Yes, Flux 2 requires a lot of system RAM and performs around three times slower than Flux Korea.
What versions does the creator provide access to?
-The creator has shared an FP8 quantized version of Flux 2 that canFlux 2 vs Qwenedit run on lower-VRAM GPUs like the 4090.
Where can viewers find the shared workflow and models?
-They are available on the creator's Patreon, which is free to join.
What future plans does the creator mention?
-The creator plans to train Loras for the Flux 2 image generator, update their Lora training tool within a week or two, and continue improving results.
Is there a community space for users to connect?
-Yes, the creator is launching a free Discord for chatting, learning, and sharing feedback.
What results did the creator get during the live test?
-The model produced a decent image with a simple prompt, showing promise even without Loras or detailed settings.
Outlines
- 00:00
🎉 Exciting News: Flux 2 is Here!
In this opening paragraph, the speaker shares their excitement about the release of Flux 2, a tool they have been following since the launch of Flux Korea. They mention how Flux 2 seems to focus on hyper-realism and might rival other tools like Quenedit. The speaker briefly discusses their experience using Flux 2 for the first time, noting its impressive improvements. They also mention uploadingFlux 2 overview a workflow example for lower VRAM users on their Patreon and the availability of a quantized version of Flux 2 for cards like the 4090.
💻 First Impressions: Great Quality, But Challenges Ahead
The speaker continues by discussing their first attempts with Flux 2, emphasizing how impressed they are with the image quality, which includes multi-angle shots. However, they also mention facing some initial issues, such as their computer crashing and the high RAM usage. Despite these challenges, they remain optimistic, highlighting that the workflow and models are available for testing on their Patreon. They also acknowledge that their system might need more optimization for smooth performance, and note the model runs slower compared to Flux Korea.
🖼️ Initial Results: Good Promise, Room for Improvement
In this paragraph, the speaker evaluates the results from the test render, acknowledging that while the outcome isn’t perfect, it holds a lot of promise. They point out that the model they used was simple and the prompt was small, so they expect more refined results with better inputs. The speaker encourages viewers to try Flux 2 themselves, as more skilled users have achieved better results. They emphasize that this video is more of an announcement and a starting point, with future improvements on the way.
🛠️ Future Plans: Training and Updates on the Horizon
The speaker reveals their plans to improve their work with Flux 2, including training specific models (Loras) and refining their workflow. They commit to updating their Laura training tool for Flux 2 in the coming weeks. They also mention that the tool and resources are available to their Patreon supporters. The speaker expresses gratitude for the feedback they've received so far and teases the upcoming launch of a community Discord, which will be free for anyone to join and engage with other Flux users.
Mindmap
Keywords
💡Flux 2
Flux 2 is the new AI image model announced and discussed throughout the video. It’s presented as an improved successor that emphasizes hyper-realism and advanced editing features; the speaker repeatedly says they woke up to its release and have been testing it for a few minutes. In the script Flux 2 is the central subject — the creator compares its speed and quality to earlier Flux models and to competitors, and describes practical details like memory needs and crash behaviour while running Flux 2 workflows.
💡Qwen Edit
Qwen Edit (referred to in the title as a competitor) is another AI system for image editing that Flux 2 is compared against. The video frames Flux 2 as a potential rival by saying Flux 2 will 'be a competitor to Qwen Edit' and highlights similar features such as using reference photos and editing existing images. Mentioning Qwen Edit helps viewers understand the market positioning: the speaker is asking implicitly whether Flux 2 is 'better than Qwen Edit' in quality and functionality.
💡Hyper-realism
Hyper-realismFlux 2 vs Qwen Edit describes the visual quality target of Flux 2 — images that look extremely realistic and detailed, often indistinguishable from high-resolution photographs. The speaker emphasizes that 'Flux 2 looks like it is very strong on hyper realism,' which explains why they find the results 'absolutely incredible' and are excited to tinker more. In context, hyper-realism is the key selling point that makes Flux 2 stand out versus older models or competitors.
💡Image editing
Image editing here refers to modifying or enhancing images using the AI model — including changing angles, expanding shots, or refining details based on prompts and reference images. The creator points out that Flux 2 allows you to 'edit images and use reference photos' much like users expect from Qwen Edit, and demonstrates examples such as expanding shots and producing multi-angle renders. In the video the workflow, crashes, and quality comparisons all revolve around the author’s experience editing images with Flux 2.
💡Reference photos
Reference photos are existing pictures supplied to the model to guide edits or ensure consistency (for example, matching a subject’s face or an object’s appearance). The speaker notes that Flux 2 supports the use of reference photos for editing, a capability that aligns it with competing tools and makes edits more controllable. In the script the author mentions using reference material and getting 'multi-angle' outputs and expanded shots, which relies on the model interpreting those references.
💡FP8 quantized version
The FP8 quantized version refers to a model variant that uses lower-precision (FP8) number formats to reduce memory usage, enabling the model to run on GPUs with less VRAM. The speaker says they uploaded an 'FP8 quantized version of Flux 2' to Patreon so people with cards like the 4090 can run it, explaining that the full model otherwise needs very high VRAM. This term is important because it explains one practical approach the author used to make Flux 2 more accessible to users without the newest or largest GPUs.
💡VRAM / GPUs (5090, 4090)
VRAM (video RAM) is the graphics memory on a GPU; the video repeatedly references specific cards (a 5090 with 32GB, and 4090 compatibility) to explain hardware requirements. The speaker explains they’re running Flux 2 on a 5090 with 32GB and that the FP8 quantized version helps users with a 4090 run the model, while noting you still likely need around 24GB. These hardware references contextualize performance observations (like crashes and seconds per iteration) and give viewers practical setup expectations.
💡Loras (LoRA)
Loras (short for Low-Rank Adaptation) are lightweight fine-tuning files or adapters used to train or adapt models to specific styles, subjects, or parameter tweaks without retraining the whole model. The speaker says they will 'get some Loras trained on this' and will update their 'Lora training tool' for Flux 2, indicating plans to create custom style or subject adapters to improve or specialize results. In the script Loras are presented as the next step for customization once the creator figures out the best parameters for Flux 2.
💡Workflow
Workflow refers to the sequence of steps, scripts, and settings the creator uses to run Flux 2 — from loading the model to generating images and applying Loras. The speaker uploaded an example workflow to Patreon and repeatedly returns to it while trying to get things running, describing crashes, RAM usage, and iteration speed as part of the workflow experience. The workflow is crucial for viewers who want to reproduce the demonstrated results or run Flux 2 on different hardware.
💡Patreon
Patreon is the membership platform the creator uses to share downloads, workflows, and extra resources with their audience. The speaker mentions uploading the example workflow and the FP8 quantized model to Patreon and invites viewers to join (noting it’s free to join) to access materials and give feedback. Patreon functions here as the distribution and community hub where viewers can obtain the files referenced in the video and interact with the creator.
💡System RAM
System RAM (computer memory) is called out as an important requirement for loading and running the Flux 2 model; the speaker warns 'you're going to need a lot of system RAM to load this model.' They show the model consuming a large portion of RAM while rendering and state that this contributed to performance problems and a crash. Mentioning system RAM helps viewers understand the full hardware profile needed beyond just GPU VRAM.
💡Seconds per iteration / performance
Seconds per iteration is a measure of how long the model takes to produce one pass or sample; the speaker reports about '6.5 seconds per iteration' and says that’s 'about three times slower than Flux Korea.' This performance metric is used in the video to quantify rendering speed and to compare Flux 2 with prior models, helping viewers weigh tradeoffs between higher image quality (hyper-realism) and slower generation times. Performance details inform viewers whether Flux 2 fits their time and hardware constraints.
Highlights
Flux 2 has been officially released, bringing major improvements in hyper-realism.
The creator expresses excitement, noting Flux was the model that initially inspired their AI journey.
Flux 2 introduces strong image-editing capabilities, becoming a direct competitor to Qwen Edit.
Users can edit images, use reference photos, and expand shots with improved quality.
The creator tested the model for only five minutes and was already impressed with results.
An FP8 example workflow is available on their Patreon to help run Flux 2 on lower VRAM.
Flux 2 still requires significant resources—at least 24GB VRAM recommended.
Testing was performed on an RTX 5090 with 32GB VRAM, though some crashes occurred.
An FP8-quantized version is shared for users with GPUs like the RTX 4090.
The model demonstrates highly impressive multi-angle and shot-expansion capabilities.
Rendering speed is about 6.5 seconds per iteration—roughly 3× slower than Flux Korea.
Early sample results show strong potential even with simple prompts and no LoJSON code correctionRAs.
More skilled users have already produced impressive images with the new model.
The creator plans to experiment further and produce higher-quality results soon.
LoRA training support for Flux 2 will be added to the creator’s training tool within one to two weeks.
All models and workflows mentioned are linked on Patreon for free access.
A new community Discord will open soon for free discussion and learning.